The Heterogeneous Effects of Eco-labels on
Internalities and Externalities
Anshuman Sahoo∗
Graduate School of Business
Steyer-Taylor Center for Energy Policy and Finance
Stanford University
and
Nik Sawe
Emmett Interdisciplinary Program in Environment and Resources
Stanford University
March 22, 2015
∗
Contact information:
A. Sahoo (corresponding author)
655 Knight Management Center, Graduate School of Business, Stanford University, Stanford, CA 94305.
Phone: +1 713 305 2204; e-mail: asahoo@stanford.edu.
N. Sawe
Yang and Yamazaki Building, Suite 226, Stanford University, Stanford, CA 94305. Phone: +1 650 814 4648;
e-mail: sawe@stanford.edu.
A previous version of this paper was circulated under the title “Negative Dividends: Internality Losses
can Outweigh Externality Gains”. We gratefully acknowledge the financial support of the Precourt Energy
Efficiency Center at Stanford University and John Beshears, Sebastien Houde, Brian Knutson, Sam McClure,
and James Sweeney for their comments on the experimental procedure and analysis. We also thank Taurean
Butler and Natalie Luu for their research assistance. All errors are our own.
The Heterogeneous Effects of Eco-labels on Internalities and Externalities
Abstract:
The Energy Star program labels energy efficient goods and has been credited with reducing
the external costs of energy consumption. Its social value is nonetheless ambiguous if, in
its absence, some consumers over-value while others under-value the energy consumption
attribute of goods, relative to their economic preferences. The label must perform opposite
tasks to guarantee an increase in consumer welfare: it must prompt some individuals to
increase their valuation of the energy consumption attribute and others, to decrease it.
Otherwise, the program could yield “negative dividends” by inducing losses in individuallevel welfare that outweigh externality reductions. We develop a method to quantify the
impact of the program on individual-level decision-making behavior and welfare. Using
novel data from a stated choice experiment involving light bulbs, we illustrate the potential
for negative dividends and that the value of programs such as the Energy Star depends on
the choice set available to consumers.
Keywords: Energy Star, Energy efficiency, Internality, Stated choice, Light bulbs.
JEL codes: D12, H23, Q48
1
Introduction
Increases in energy efficiency are often considered among the most cost-effective methods
to reduce energy demand and the environmental externalities associated with energy consumption. Since residential use accounts for about 22% of U.S. energy consumption, changes
in appliance stock could meaningfully shift U.S. energy demand. However, consumers frequently choose not to purchase energy efficient alternatives that should provide them with
net savings. In seminal work, Hausman (1979) showed that consumers choosing appliances
do not buy goods that would have been optimal at a market interest rate but act instead as
if they discount future savings on energy consumption at a much higher rate. Since then,
a large body of literature has attempted to explain how consumers make such decisions
(Gillingham and Palmer, 2014).
Embedded in several explanations is a notion that consumers do not pay enough attention
to the energy consumption attribute of alternative appliances.1 Standard economic theory
assumes that all consumers make consumption decisions as if they were maximizing a utility
function using all available information; the rational consumer would thus pay full attention
to the energy consumption attribute. Nonetheless, inattentiveness may be pervasive in consumer choice. For example, Chetty, Looney, and Kroft (2009) show that consumers imperfectly consider taxes on common consumer goods. In the energy sector, Allcott (2011) shows
that U.S. consumers do not pay adequate attention to fuel costs in making vehicle purchase
decisions; Leard (2013) similarly suggests that approximately 30% of consumers shopping
for a new vehicle completely ignore fuel costs. In the context of appliance choice, Houde
(2014) presents evidence characterizing three types of appliance consumers, two of which
effectively ignore electricity cost implications. If consumers are insufficiently attentive to the
ramifications of energy consumption for themselves, they would certainly be insufficiently
attentive from a social perspective, which adds the external costs of energy consumption.
Thus, given an expected pattern of product use, the under-investment in energy efficiency
by consumers implies a greater than socially optimal level of energy consumption.2
1
We focus on an inattentiveness hypothesis because this, in our opinion, most directly motivates the
apparent goal of the Energy Star to increase the salience of the energy consumption attribute of goods. This
also motivates the approach in Sallee (2014), in which the consumer must incur significant costs to determine
the level of energy efficiency implied by alternative goods. Nonetheless, a large number of alternative
explanations have been offered for the perceived low uptake of energy efficient goods.
2
Throughout our discussion, we restrict attention to product choice decisions and assume that the pattern
1
A natural policy reaction is to increase the salience of the energy consumption attribute
of alternatives. Common policy responses implement this reaction to differing degrees. One
strategy, that of minimum efficiency standards, does not attempt to increase the salience but
instead simply removes energy inefficient goods from the market. This approach entails a cost
to certain consumers: those who plan to infrequently use the affected good may be compelled
to invest in a higher level of efficiency than they would rationally choose. A strategy that
can theoretically achieve the first-best outcome is that of attaching a Pigouvian tax on
energy-consuming goods. Such goods exert external costs attributable to environmental
damages.3 Standard theory prescribes that goods or services should be taxed at a level
equal to the marginal external costs they produce. Thus, to achieve the first-best outcome,
the policymaker would in general need to know each individual’s rate of utilization of the
affected energy-consuming good. Absent this information, we would expect a gap to exist
between the first-best policy outcome and that actually obtained.
Two strategies attempt to affect only those consumers who do not fully consider the
energy consumption attribute. One of these seeks to increase the salience of energy consumption information by explicitly providing it at the time of decision-making. Another
class seeks to do so by using “nudges,” or changes in choice architecture, as described by
Thaler and Sunstein (2008). By definition, these strategies preserve all alternatives in consumers’ choice sets but change the manner in which they are presented. Critical to the
appeal of such policies is the supposition that they should not affect the decisions of those
consumers who act rationally in the absence of the intervention.
The Energy Star certification and labeling program administered by the U.S. Environmental Protection Agency (EPA) can be interpreted as one such policy. Manufacturers of
goods that meet energy consumption benchmarks may voluntarily label compliant products with the Energy Star label. In general, such eco-labels have increased the uptake of
of use remains the same regardless of the product chosen. A study of the implications of the Energy Star
for the misoptimization of product use would provide a useful contribution. Our statement about excessive
energy consumption relative to a social optimum would be reversed only if consumers’ product use increased
so radically with goods of higher efficiency that this effect counteracted the lower expected consumption at a
constant level of use. The increase in use that accompanies greater energy efficiency is known as the rebound
effect, and most estimates of it suggest that our assertion holds (Gillingham and Palmer, 2014).
3
They also exert external costs attributable to energy security, since societies maintain military forces
partially to guarantee access to energy resources. Since this paper focuses on appliances in the U.S. context,
where a very small portion of electricity is generated from resources that are sourced internationally, we do
not focus on this second source of externalities.
2
energy efficient goods; Sammer and Wustenhagen (2006) provide evidence that consumers’
willingness-to-pay for products with energy eco-labels exceeds that for products that are
equally efficient but unlabeled. Further, Ward et al. (2011) estimate that customers are
willing to pay an extra $250 - $350 for a refrigerator with an Energy Star label.
If the analyst’s policy goal is solely to reduce environmental externalities, this evidence
suggests that the Energy Star has been effective. The expression by consumers, on average,
of a higher willingness-to-pay for labeled goods implies that they will more frequently select
energy efficient alternatives. Nonetheless, this evidence does not comment on whether the
program also makes individuals better off, as measured by the degree to which it encourages
them to choose products in a manner consistent with utility maximization.
Indeed, while consumers may be willing to pay for eco-labels, it is unclear whether the
latter increase or decrease the former’s attention on the energy consumption attribute of alternatives. This bidirectionality may underpin recent work showing that consumers respond
heterogeneously to eco-labels such as the Energy Star. Both Houde (2014) and Shen and
Saijo (2009) suggest that consumers may (1) use labels as a substitute for conventional utility
optimization and ignore electricity cost and consumption data, (2) use electricity cost and
consumption data to optimize choice but ignore the eco-label, or (3) use neither electricity
cost and consumption data nor eco-labels in making a choice. These observations imply
a second level of uncertainty. After the Energy Star exerts an impact on the attention an
individual pays to the energy consumption attribute, he or she may make choices that reflect
a lower or higher valuation of savings on energy consumption, relative to that expressed in
the absence of the label. An implication is that some individuals who considered energy
consumption rationally in the absence of the label may use the attribute irrationally or not
at all in the presence of the Energy Star. If such changes occur, one of the chief appeals of
the label and similar behavioral instruments would be at risk.
More broadly, the possibility that the Energy Star could increase some consumers’ valuation on energy consumption savings and decrease it for others raises the question that
motivates this paper: what is the value of the Energy Star to consumers? We ask this question because the heterogeneous effects on consumers’ valuation suggest that the Energy Star
could sometimes influence consumers to make decisions that are less consistent with utility
maximization. This paper presents results from a stated choice experiment on light bulb
choice intended to answer our question. In the experiment, a cross-national sample of 1,550
3
individuals made hypothetical choices across compact fluorescent light (CFL) bulbs. Using
the 46,500 choices we observe in these data, we estimate coefficients on a mixed logit specification stemming from an underlying random utility model. Importantly, the estimation
procedure yields utility coefficients, conditional on the sequence of an individual’s choice, on
both the energy consumption attribute of alternatives and the interaction between it and
the presence of the Energy Star. We thus estimate each individual’s utility coefficient on
energy consumption in the absence and presence of the Energy Star.
This allows us to include in the evaluation of the Energy Star program its impact not only
on environmental externalities but also on consumers’ internalities, as defined by Herrnstein
et al. (1993) and as applied by, e.g., Allcott, Mullainathan, and Taubinsky (2014) and
Leard (2013). We expand considerably on the concept of internalities subsequently, but
they can be understood as losses in experienced utility, relative to a rational benchmark.
Consider a given consumer at different times as different people. This is consistent with
standard economic theory as long as the consumer assigns equal weights to all versions of
herself in her decision-making process, as this implies that her choices reflect time-consistent
preferences. However, if the consumer weighs different versions of herself unequally such that
the experienced utility implications of a choice at different times are scaled unequally, her
choices reflect time-inconsistent preferences. These choices could imply a loss in experienced
utility, and we term such a loss as an “internality.” If an under- or over-valuation of energy
consumption savings influences a shift in the good selected away from that consistent with
the consumer’s true preferences, these tendencies will exert internalities on the consumer. If
the Energy Star affects the valuation of energy consumption savings, it implies changes in the
magnitude of internalities. Our central point in this paper is that these must be quantified
in economic analyses of the Energy Star and, in general, of similar behavioral tools.
We facilitate this quantification by establishing a rational benchmark utility coefficient on
energy consumption for each individual; this can be interpreted as reflecting time-consistent
preferences for energy consumption savings. The benchmark follows from a discount rate
elicited from each individual and assumptions about the value of savings on energy consumption. When the consumer acts as if she is applying this coefficient in decision-making, she
chooses the good that maximizes her experienced utility. Since a wedge between the rational
benchmark utility coefficient and that observed in either the absence or presence of the Energy Star could prompt the selection of a good that does not maximize experienced utility, it
4
could yield an internality. The larger this gap, the larger the potential for misoptimization
of product choice and of losses in experienced utility to the consumer. Upon estimating the
size of the gaps in the presence and absence of the Energy Star, we can determine whether
the label mitigates or exacerbates such wedges at both the individual and population level.
The chief conceptual contribution of this paper is a method by which to value the impacts of behavioral policy instruments, including their implications for both externalities
and internalities. The latter usually have not been formally included in such assessments.
While our methodological approach is motivated and illustrated by the Energy Star, it can
be applied to other contexts as long as the analyst is willing to make two assumptions. First,
aside from the utility coefficient on the attribute(s) affected by the behavioral intervention,
all estimated utility coefficients reflect time-consistent preferences. The second assumption
is that of the nature of long-term costs and benefits. In contexts such as ours, in which individuals face an immediate cost and long-term benefits or an immediate benefit and long-term
costs, the second assumption requires assumptions about the magnitude of these benefits or
costs and the duration over which these obtain. The specification of these costs and benefits
becomes much more difficult in scenarios where estimates of the experienced utility losses
over time reflect much more heavily the analyst’s judgment. Importantly, we highlight that
the implications of behavioral instruments for internalities and externalities are functions of
the choice set available to the consumer. Some choice sets may be characterized by a dominant alternative such that, even when wedges exist between the rational benchmark value
of an utility coefficient and that actually used in decision-making, no changes in experienced
utility (i.e., internalities) obtain. This is in contrast to those choice sets offering a rich set
of trade-offs to the consumer. In such settings, the experienced utility realized is sensitive
to the value of the utility coefficient.
Our main findings in the Energy Star context are as follows. In aggregate, the Energy
Star label appears to do what it was designed to do: it increases consumers’ valuation of
savings on energy consumption. More often than not (for 53% of our participants), the effect
of the Energy Star is aligned with the individual’s own interest. The overall value of changes
in experienced utility across the population and attributable to the program depends on the
marginal utilities of wealth among the affected population. We examine the performance
of the Energy Star on a choice set offered by a national retailer and find that the program
yields modest increases in experienced utility and decreases in external costs. We observe
5
that the overall value of the Energy Star program derives principally from its impact on
internalities instead of on externalities. While we do not present results from an alternate
choice set, this positive result for both internalities and externalities is not a general result
across all choice sets.4
The work in this paper contributes to two bodies of literature. One studies decisionmaking processes in appliance choice. Our work is particularly motivated by Houde (2014) in
quantifying the implications of heterogeneous responses to the Energy Star that he identifies.
It also extends the study of eco-labeling, which has hitherto focused mostly on measuring
consumers’ willingness-to-pay (Sammer and Wustenhagen, 2006; Shen and Saijo, 2009; Ward
et al., 2011). The work by Newell and Siikamaki (2014) considers how the informational
content of labels guide decision-making. While we overlap conceptually with them in using
an objectively determined optimal behavior construct, we differ in focusing on individual
differences in response to the Energy Star label rather than on differences in response to
alternative labels. Like us, Min et al. (2014) focus on the Energy Star, but they present
aggregate-level implications of eco-labeling on consumers’ decision-making processes.
The other literature studies how the interactions of internalities and externalities change
the choice and use of policy instruments. Our work follows Leard (2013) closely in its
treatment of internalities but differs in both the context and the examination of a nonprice policy instrument. More broadly, our work is related to Allcott, Mullainathan, and
Taubinsky (2014), who consider how the presence of both externalities and internalities affect
energy policy-making. They suggest that pecuniary instruments to offset externalities yield
a double dividend by also reducing internalities. However, they suggest that if such taxes
are poorly targeted, other policies, such as information provision and nudges may be more
appealing. Their intuition is that these policies are designed to affect only those who do not
act rationally. Our work suggests an important nuance: the heterogeneous effects of such
behavioral instruments may imply a negative dividend.
Although our paper illustrates this nuance in the context of policy evaluation, analogous
lessons apply to managers. While changes in choice architecture could increase profits,
certain classes of consumers may preferentially gain while others lose. Though regulatory
guidance on choice architecture is unclear, if the harm to the latter segment of consumers is
sufficiently large, changes in choice architecture may prompt action by bodies with a mandate
4
We provide a specific counter-example in Section 5.3.
6
to protect consumer welfare. More directly, managers should not assume that the average
change in consumer response following a change in choice architecture will apply to an entire
population of consumers. Our illustration in the eco-labeling context provides one example of
the importance of individual-level differences in response behavior. While average changes in
consumer behavior may imply an increase in profitability, the distribution of responses may
suggest just the opposite. Managers should attempt to characterize individual differences in
consumer behavior to ensure that they both avoid reprimands by consumer watchdogs and
select profit maximizing strategies.
2
Experimental and Analytical Methods
Our analysis uses compact fluorescent light bulb (CFL) choice data from a stated choice
experiment. In this section, we describe our experiment and economic model of choice.
2.1
Experimental Details
Figure 1 describes the experimental procedure. We contracted with a third-party survey
administration firm to recruit a cross-national panel. Eligible participants were U.S. residents
who were at least 18 years old, owned their primary residence, and had either purchased or
remodeled their residences within five years. The last criterion sought to limit participants
to those who had recently made decisions about energy-consuming goods.
Figure 1: Overview of experimental flow
Those participants who did not meet the inclusion criteria or incorrectly answered one
of two questions designed to test attentiveness were dismissed without compensation, while
7
those who did were randomized into one of three conditions.5 This variation was a precursor
to an experiment that will be implemented separately. The three conditions differed in the
denomination of the consumption information provided. We re-phrased the accompanying
interpretive information to match the denominations. Condition 1 provided participants
with information in only power (Watt (W)) units. In Condition 2, participants received information in terms of the estimated annual cost of operating the good. Condition 3 provided
consumption information in both units. The Appendix provides the interpretive information accompanying each condition. We adjust the data to account for differences in error
variances across conditions (see Section 3.1).
Participants performed a series of ten light bulb and refrigerator choice tasks. This paper
focuses on the light bulb data to isolate our message about the implications of heterogeneous
responses to the Energy Star for the value of the program. For each product, twenty choice
sets were organized into two blocks of ten choice sets. Following the randomization step,
individuals responded to randomly assigned light bulb and refrigerator choice blocks. We
generated the constituent choice sets using Ngene (ChoiceMetrics, 2012). Our program
applied priors on coefficient values from a pilot stated choice experiment to build attribute
levels designed to optimize the efficiency of our estimation. The range of prices and energy
consumption were such that all alternatives could have been Energy Star certified.
As Figure 2 illustrates, each participant selected one of three alternatives from each of
ten choice sets. An alternative is characterized by three primary attributes: price, energy
consumption, and the Energy Star certification. To increase the reality of the task, we labeled
each bulb as providing light of one of four colors.6 Each participant received instructions
about the task, information with which to interpret the attributes, and a set of assumptions
to make about attributes, such as the lifetime of the light bulb, which we did not vary.
The Appendix details this material. We collected 46,500 choice observations from our 1,550
participants.
Subsequently, each participant answered 144 demographic, psychological, and financial
questions intended to characterize individuals.7 We consolidated many of these questions
into standard metrics (e.g., 21 questions about time preferences implied an estimate of the
5
As we discuss later, we did not find material differences among consumers’ characteristics or choice
behaviors across conditions.
6
The color describes the warmth of the light as “warm white,” “soft white,” “neutral” or “cool.”
7
The Appendix provides details about these questions.
8
Figure 2: Example bulb choice set
participant’s temporal discount rate). Following this consolidation, our data included 35
demographic, psychological and financial variables.
2.2
2.2.1
Economic Model
Base utility assumptions and choice model
We assume that participants choose the alternative yielding the highest decision utility and
thereby model choices with the random utility model (RUM). Formally, individual i selects
good j among J alternatives in set t of T choice situations to maximize her utility:
Uijt = Vijt + εijt = βXijt + εijt
(1)
Xijt is a vector of explanatory variables that includes the attributes of alternatives, Vijt
or βXijt is the deterministic component of utility, and εijt is a random error term. The
probability, Pijt , that individual i selects alternative j over k in choice set t is:
Pijt = P (Uijt > Uikt )∀k 6= j ∈ t → Pijt = P (εikt − εijt < βXijt − βXikt )∀k 6= j ∈ t
(2)
A traditional assumption is that the error terms are identically and independently distributed
(iid) according to an extreme value type I distribution. With this assumption, we obtain
9
the multinomial logit (MNL) model, which implies that the probability that alternative j is
selected, conditional on β, is given by:
eσβXijt
Pijt (β) = PK
σβXikt
k=1 e
(3)
Above, σ is a scale parameter on utility that enters the model from our assumed joint
distribution of εijt . The distribution is characterized by both location and scale parameters.
We return to the scale parameter in Section 3.1 but assume now that it has been normalized
to 1.
The MNL model has three limitations that we overcome by estimating a mixed logit (ML)
model. ML models allow for random taste variation, unrestricted substitution patterns by
permitting any specification of the correlation pattern across alternatives, and a correlation
in unobserved factors over choices made by the same individual (Revelt and Train, 1998).
2.2.2
Allowing preference heterogeneity through the mixed logit model
Mixed logit models “mix” MNL estimates by integrating over a joint density of parameters
characterizing distributions of coefficients on particular attributes of the alternatives (Train,
2009). As such, we update the probability that alternative j is selected in choice set t to:
Z
Pijt =
eσβi Xijt
PK σβ X f (β | Ω)dβ
i ikt
k=1 e
(4)
Equation 4 introduces a mixing distribution, f (β | Ω), over all random coefficients.
The set Ω includes parameters of the distributions on random coefficients. If one specifies
that marginal coefficient distributions are correlated, as we do, the parameter space also
includes co-variance terms between correlated coefficients. Since an analytical solution for
the likelihood function does not exist, we estimate coefficients by maximizing a simulated
log likelihood function.8 The function is defined by the product of the probabilities that
individual i purchases the product actually chosen, j ∗ , in choice situation t:
SLL =
XX
t
8
lnPij ∗ t
(5)
i
Train (2009) and Hensher and Greene (2003) provide additional details about simulated maximum
likelihood.
10
Briefly, the procedure assumes parameter values for all elements of Ω, draws coefficient vectors, given these parameter values, calculates the probability of selection of the alternative
actually selected, given the coefficient vector, repeats the first three steps many times, averages the probability implied by each of these repeats, and selects the parameter values
ΩSM LE that yield the highest simulated likelihood.9
2.2.3
From the mixed logit to “individual-level” parameters
Since we observe repeated choices from each participant, we can refine the mixed logit-based
coefficient estimates by conditioning on the sequence of choices, yei , made by participant i
in the sequence of choice situations, tei . We accordingly assign each individual a conditional
parameter distribution, rather than the sample-wide parameter distribution. Following Train
(2009), upon conditioning on tei and yei , we update the distribution of coefficients based on
ΩSM LE , denoted g(β | ΩSM LE ), to h(β | yei , tei , ΩSM LE ).
We do not provide the full intuition behind the distribution h, since it is available in
Train (2009). However, h re-scales the distribution g, as shown in Equation 6:
P (yei | tei , β)g(β | ΩSM LE )
h(β | yei , tei , ΩSM LE ) =
P (yei | tei , ΩSM LE )
(6)
Equation 6 follows from an application of Bayes’ Rule and indicates that the conditional
density of the coefficient vector, h, among those who choose the sequence yei when choosing
among tei is proportional to the product of the unconditional density and the probability that
yei would be chosen if the coefficient vector were β.
We estimate the conditional distribution of coefficients for each individual by simulation.
The simulation draws coefficient values β from the population density, g(β | ΩSM LE ), calculates the probability of observing yei given this draw, and takes the weighted average of the
draws. The weights are determined by the ratio of the calculated probability for a particular
draw to the sum of probabilities across all draws. (Train, 2009)
Since ΩSM LE is itself estimated with sampling error, we add a second layer of simulation.
This second layer draws Ω from N (ΩSM LE , WSM LE ), the estimated sampling distribution of
Ω, using a Choleski decomposition of WSM LE (Train, 2009). Thus, the overall simulation
procedure entails multiple draws of both β and Ω. We accomplish the latter by taking 500
Halton draws and the former by taking 500 samples from the sampling distribution.
9
The subscript “SMLE” denotes simulated maximum likelihood estimator.
11
Though we refer colloquially to the resulting βi as individual-level coefficients, they are
correctly interpreted as population coefficients, conditional on the choices made by individual i. Since the data generation process includes a fixed number of observations for each
participant, the conditional mean coefficient vector, β̄i , is not a consistent estimate of βi .
Finally, the individual-level coefficients imply individual-level WTP measures on attribute l of alternatives:
βil
(7)
ηi
Above, ηi is the marginal utility of wealth. The individual-level measures imply a sampleW T Pil = −
wide mean and standard deviation of the WTP.10
2.2.4
Specifying the observed component of utility
We have so far abstracted from the observable components of utility, which Equation 8
details:11
Vij = ηi · pj + θi · Cj + λi · ESj + γi · (ESj )(Cj ) + ΥTi Zj
(8)
In the above model, pj is the price of product j, θi is the marginal utility of an additional
unit of power rating, Cj is the power rating of the good, λi is the marginal utility from
the presence of the Energy Star logo, ESj is a dummy variable equal to 1 if the product is
Energy Star labeled, and γi is the incremental marginal utility of an additional unit of energy
consumption when the product is Energy Star labeled. Finally, Υi is a vector of marginal
utilities from the warmth of the light emitted, and Zj is a vector of dummy variables on
the warmth levels. Though we allow the utility to reflect the color of light produced by a
bulb, we assume that the utility from the lighting service itself is uniform across all bulbs
and irrelevant to our modeling of differences in utility. Given a particular pattern of product
use, the energy consumption of the bulb is a monotone transformation of the power rating.
10
We also compute parameters of the WTP distribution by performing 1,000,000 draws of the constituent
mixed logit utility coefficients. Though we do not report the simulation-based WTP measures, the mean
and median WTP measures are close when calculated by either approach. The standard deviations from the
simulation-based method are much larger than those implied from the aggregation of the individual-level
coefficients. This reflects the larger variance in the unconditional distribution relative to the conditional
distribution.
11
As mentioned above, this model should be interpreted as describing decision utility instead of experienced
utility.
12
Since our experiment entailed specific assumptions about the usage pattern, we refer to the
attribute as energy consumption.12 Note that this specification allows utility to increase or
decrease by the presence of the Energy Star label independently of the energy consumption
of the particular product. This affords us the ability to separately assess whether the Energy
Star label leads people to place more or less weight on the actual energy consumption.
We allow four utility coefficients to vary randomly across individuals: ηi , θi , λi , and γi .
Since theory suggests that utility decreases in prices and energy consumption, we constrain
ηi and θi to be strictly negative by using a lognormal mixing distribution for both.13 Theory
does not provide such guidance for the other parameters, and we use a normal mixing
distribution for them. Since heterogeneity in preferences for light bulb warmth levels is
uninteresting in this context, we set Υi = Υ ∀ i. Finally, we allow the coefficient distributions
to be correlated, since theory suggests, for example, that individuals with high marginal
utilities of wealth may also have high marginal disutilities from energy consumption.
3
Aggregate Results
5,919 respondents entered our experiment, and we arrived at our sample of 1,550 upon
removing those that either did not meet the inclusion criteria or correctly respond to the
attentiveness questions. We present summary demographic, psychological, and financial data
in Table 6 of the Appendix; the lack of significant differences across the three conditions
suggests that our sample is balanced across them. In Section 3.1, we provide details about
our estimation procedure, and in Section 3.2, we discuss our estimated ML coefficients.
3.1
Estimation Details
Three comments about our estimation procedure clarify the interpretation of the coefficients.
First, regardless of the units of consumption with which a participant was presented, a deterministic link between bulb wattage and expected annual expenditure on electricity allows us
to use a singly denominated vector for estimation. Second, the standard errors we use for inference are based on cluster robust variance estimators (CRVE), with observations clustered
12
Moreover, consumption, not the rating, is the source of the internalities and externalities we subsequently
consider.
13
A lognormal distribution for price moreover assures finite moments for willingness-to-pay distributions
(Daly, Hess, and Train, 2012).
13
at the individual level. This allows for heterogeneity in the error structure across individuals
and for an arbitrary autocorrelation structure in the errors across choice situations faced
by the same individual (Wooldridge, 2010). Finally, we relax the assumption that the scale
parameter in Equation 3 equals one. This assumption generally allows one to estimate β
even though it is not separately identified from σ by the data. Since we combine data from
three experimental conditions, we need to ensure that our estimate of β is not influenced
by differences in the error variance across conditions. To adjust for possible differences, we
follow methods developed by Swait and Louviere (1993). If the hypothesis of equal scale
factors across conditions is rejected, the method estimates condition-specific scale factors
with which to adjust the co-variate matrix prior to estimation. The Appendix details our
derivation of scale parameters on Conditions 1 and 2. The results we present here are based
on estimations with condition-specific scale parameters, but they do not change appreciably
if we reset them to equal 1.
3.2
Mixed Logit Coefficient Estimates
Table 1 presents the results of our mixed logit estimation.14 Since the ML estimates follow
from a model with dummy terms on three of the four warmth levels, the outside option is a
bulb of the remaining warmth level, of average price and consumption, and without Energy
Star certification. Coefficients should thus be interpreted as conditional on bulb purchase.
The baseline bulb can be of any of the four warmth levels, as the coefficients on price, energy
consumption, Energy Star, and the interaction term average over all warmth levels. Column
3 provides the means from the average of conditional (individual-level) coefficients.15 The
agreement between Columns 2 and 3 provides a check that the model is correctly specified
14
It is common to comment on whether the signs matched expectations. Though the means of the coeffi-
cient distributions in Column 2 imply that the signs on price and consumption are as expected, we note that
these simply follow the assumptions of our mixing distribution. However, we observe the same signs in a
model with normal mixing distributions on these attributes. The notion of an ‘expected sign’ is ambiguous
for bulb warmth levels, for the Energy Star label, and for the interaction between the Energy Star and
consumption level. The positive coefficient on the Energy Star logo is expected to the extent that we assume
consumers interpret this as some indicator of quality. We use the Stata ‘mixlogit’ command written by Hole
(2007) for our estimation.
15
We modify Hole’s ‘mixlbeta’ command to estimate individual-level means and variance-covariance matrices.
14
and accurately estimated (Allenby and Rossi, 1998; Train, 2009).16
Dependent variable: Alternative j selected by person i (binary outcome)
Attribute
Mean
-0.684
Price
(0.045)
1.839∗∗∗
Energy Star
(0.093)
Consumption
-1.056∗∗∗
(0.061)
-0.126∗∗∗
ES*Cons.
(0.026)
0.186∗∗∗
Soft white
(0.041)
-0.189∗∗∗
Neutral
(0.054)
-0.743∗∗∗
Cool
(0.060)
LogLik
Mean (IL)
∗∗∗
-1.277
1.837
Variance
1.825
(0.129)
2.403∗∗∗
(0.229)
1.48∗∗∗
-0.734
-0.126
WTP, Mean
WTP, Med.
WTP, SD
–
–
–
$4.97
$2.36
$6.46
-$1.23
-$0.71
$1.65
-$0.15
-$0.13
$0.31
∗∗∗
(0.149)
0.019∗∗∗
(0.006)
–
–
$0.51
$0.35
$0.48
–
–
-$0.51
-$0.35
$0.49
–
–
-$2.02
-$1.39
$1.93
-11914.764
Obs.
46,500
Correctly
Predicted (%)
60.2
Table 1: Estimated ML parameters for bulb choice. Columns 5 through 7 list summary
statistics on WTP measures. Numbers in parentheses are standard errors. Note: ES∗Cons.
indicates the Energy Star and Consumption interaction term. Key to statistical significance
–
∗∗∗
: ≤ 0.001;
∗∗
: ≤ 0.01; ∗ : ≤ 0.05.
The variance terms in the fourth column validate our consideration of heterogeneous
impacts of the Energy Star. All terms are statistically different from zero, and this confirms
significant heterogeneity across all four coefficients. By considering this heterogeneity, the
ML model outperforms the MNL in correctly predicting 60.2% of choices relative to the
latter’s 51.6%.17
We subsequently consider the implications of heterogeneity, but our aggregate measures
imply that participants express a positive WTP for Energy Star certification independently
of actual energy consumption. To put the WTP for the Energy Star label in context, the
16
The mean coefficients on price and consumption are parameters of the lognormal distribution; the
corresponding parameters of the normal distribution are -1.257 and -0.728, respectively.
17
For individual i and choice set t, the model correctly predicts choice if it assigns the highest probability
of selection to j ∗ , the alternative actually selected.
15
mean price of bulbs in the study was $5.56. Further, certification decreases the marginal
utility from a 1W higher power rating; thus, Energy Star certification makes consumers
more sensitive on average to the energy use of the bulb. In our upcoming discussion, we
will refer interchangeably to the effect of the Energy Star on the marginal disutility of a 1W
higher power rating and on the marginal utility of energy consumption savings implied. If
the Energy Star increases the former, as on aggregate, it also increases the latter.
The estimate of the coefficient on the Energy Star label itself implies that the certification
endows products with a “brand” that is almost always positively valued.18 The magnitude
of the median willingness-to-pay for certification, $2.36, is moreover greater than the overall
incremental willingness-to-pay for a one watt lower power rating by a factor of approximately
three.19 The label exerts a larger impact on light bulb choice by imparting a “brand name”
than by affecting consumers’ sensitivity to energy consumption information. To the extent
that the Energy Star label actually distinguishes efficient light bulbs in the market from
inefficient ones, this energy-independent Energy Star effect yields additional reductions in
the external costs of electricity generation and consumption.20 However, as we make clear
in the subsequent section, there are instances in which an individual is compelled by the
Energy Star program to be overly attentive to energy consumption, relative to an individually
rational level. In that case, the additional impact of this energy-independent Energy Star
effect would make the individual even worse off than does the impact of the label on the
utility coefficient on energy consumption. The social value of this effect is thus ambiguous.
We comment more on the role of this coefficient in Section 5.2 and illustrate its effect in our
example evaluation in Section 5.3.
Finding 1 summarizes our findings from the ML model:
Finding 1: In aggregate and for our sample, the Energy Star increases consumers’ marginal
utility from savings on energy consumption. It also increases consumers’ willingness-to-pay
for a bulb, independent of the actual energy consumption. Mixed logit estimates confirm
18
19
60 of the 1550 individual-level coefficients on the presence of the Energy Star label are negative.
The overall incremental WTP for a one watt lower power rating includes both the WTP stemming from
the incremental Energy Star effect, $0.13, and the baseline WTP of $0.71. The “brand” effect of the Energy
Star implies a median WTP that is roughly 18 times greater than the incremental WTP motivated by the
Energy Star label for a one watt lower power rating.
20
We make this qualification because it is possible that a consumer’s choice set could include bulbs of
equal efficiency with some bearing the label and others not. In this scenario, the energy-independent Energy
Star effect does not influence reductions in external costs.
16
heterogeneity in the coefficients we allow to vary.
4
Heterogeneous Impacts of the Energy Star
The significant variance of the Energy Star and Consumption interaction term in Table 1
implies that the Energy Star could either increase or decrease a participant’s marginal utility
coefficient on savings from energy consumption.21 Intuitively, increases would be valuable
to those who undervalue these savings but harmful to those who do not or who overvalue
them, relative to a rational benchmark. Section 4.1 develops this intuition by introducing
the concept of internalities and demonstrating the implied economic costs to consumers.
4.1
The Origin and Cost of Internalities
In this subsection, we consider the consumer at different time points as different people to
illustrate the origin and cost of internalities. We show how a consumer behaving as if she is
applying time inconsistent preferences to a choice situation bears an economic cost, relative
to the scenario in which she acts as if she applies the time consistent preferences that are
central to standard economic theory.
In our discussion, we rely on two notions of utility: experienced utility (Ue ) and decision
utility (Ud ). The experienced utility quantifies the benefits the consumer truly experiences
from consumption. In contrast, the decision utility quantifies the benefits she believes she will
experience from consumption at the time of decision-making. We illustrate these concepts
with both an abstract choice setting and a numerical example. Both our abstract and
numerical choice settings entail three time periods, with time progressing from t = 0 to
t = 2. At time zero, the consumer becomes aware of the choice task, which she must
complete at time one. Without loss of generality, the alternatives are such that they yield
an immediate benefit enjoyed at time one and a delayed cost experienced at time two.
At time zero, the consumer studies the choices as a fully rational decision maker. She applies her preferences to evaluate the future costs and benefits of the alternatives, x, available
in the choice set, S. She identifies the alternative that yields the largest stream of utility.
Formally, she identifies, but does not yet choose, the alternative x∗ that maximizes her experienced utility function: x∗ = argmax(x ∈ S)Ue (x). The experienced utility function, Ue ,
21
Figure 9 in the appendix presents the distribution of the coefficient estimates.
17
can be represented as in Equation 9:22
Ue (ut , ut+1 , ..., uT ) =
T
X
δ τ uτ
(9)
τ =t
Above, Ue reflects an inter-temporal preference that considers both the current instantaneous
utility, ut , and future instantaneous utility levels. δ is a parameter that measures the degree
to which utility experienced in a future period is equivalent to utility experienced in the
current period; δ assumes values between 0 and 1, and the closer it is to zero, the lower the
degree to which future utility is equivalent to current utility. We use x∗ and the experienced
utility it entails as a normative benchmark.
The consumer actually makes her consumption decision in time one, and internalities
emerge from this decision-making context if her choice at time one is not x∗ . To illustrate
this, we follow a modeling strategy from behavioral economics (O’Donoghue and Rabin, 2000;
Laibson, 1997). This introduces a parameter βu to the utility representation in Equation 9
that applies to the utility streams in all but the current time period.23 The utility function
with this new parameter is the decision utility function, Ud :
t
Ud (ut , ut+1 , ..., uT ) = δ ut + βu
T
X
δ τ uτ
(10)
τ =t+1
If βu 6= 1, the consumer acts as if she is no longer applying time-consistent preferences at the
time of decision-making. The alternative chosen at time one is xd , where xd = argmax(x ∈
S)Ud (x).
Traditionally, βu has been restricted to the range (0, 1], since Equation 10 has usually
been applied to scenarios in which consumers are characterized by preferences for immediate
gratification. In this range, when the consumer arrives at time one, she under-weights the
utility streams in future periods. Such preferences are generally known as time-inconsistent
preferences and the modeling strategy, as beta-delta discounting. Our treatment differs from
the traditional application in requiring only that βu > 0. When βu > 1, the consumer acts
as though she has a preference for delayed gratification.
Unless xd = x∗ , the optimization of the decision utility function instead of the experienced
utility function implies a loss to the decision-maker. The experienced utility implied by xd
22
23
The notation and discussion here are inspired by those in O’Donoghue and Rabin (2000).
We use the subscript ‘u’ to distinguish this parameter from the abstract coefficients of the ML model
introduced in Section 2.2.
18
is given by Ue (xd ). We refer to the gap between Ue (xd ) and Ue (x∗ ) as the internality:
Internality(xd ) = Ue (x∗ ) − Ue (xd ). The internality characterizes the loss in experienced
utility due to the use of time-inconsistent preferences. If βu < 1, the internality stems
from an under-weighing of future utility streams. If we consider the individual at times one
and two as different people, we observe that the loss in “social” value when βu < 1 stems
from the time-one consumer “overindulging” in the benefits of the consumption good such
that it imposes extra costs on the time-two individual, relative to the rational benchmark.24
Similarly, if βu > 1, the internality stems from the time-two individual limiting too greatly
the ability of the time-one individual to indulge in the benefits of the consumption good,
relative to the rational benchmark.
The Appendix includes a numerical example that illustrates these utility concepts. The
example demonstrates several points that help one understand the implications of our Energy
Star experiment. Most importantly, the example highlights that a gap between the decision
and experienced utility functions does not necessarily imply that the consumer’s choice will
change. Consequently, such a gap does not immediately signal that the consumer bears a
loss from time-inconsistency or other behavioral phenomena generating the wedge between
the two utility functions.
4.2
Measuring the Utility Implications of the Energy Star
In this subsection, we develop an intuition about the effect of the Energy Star on losses
in experienced utility and, equivalently, on internalities. We proceed here with three assumptions that we relax in our actual evaluation in the next section. First, we assume that
the light bulb choice set is continuous, or that between the minimum and maximum price
and power ratings, goods with all intermediary non-dominated combinations of prices and
power ratings are available.25 Second, we assume that consumers’ decision utility functions
do not reflect an Energy Star “brand” effect; that is, we do not discuss here the role of
the energy-independent Energy Star effect in directly influencing choice. Third, choices are
known with certainty. Though we subsequently drop these assumptions in Section 5.2, they
allow us to build a general perspective about the impact of a behavioral instrument on the
24
25
Our “society” here is the consumer comprised of both the time-one and time-two individuals.
A price and power rating combination would be dominated if, for a given power rating, a lower price
option exists. Thus, for any given power rating, the infinite choice set would include only one option at a
given price.
19
choices individuals make and on the implications for experienced utility.26
Equation 8 implies an experienced utility function for person i who has selected good j:
Ueij = ηi · pj + θi · Cj + ΥT Xj
(11)
We note that the experienced utility formulation does not include the direct “brand” effect of
the Energy Star label. This reflects a belief that it does not actually contribute to differences
in experienced utility.27 The time-consistent consumer who acts as if she were maximizing
Equation 11 selects a bulb x∗ and enjoys a utility Ue (x∗ ). Time-consistency in the light bulb
context relates to a full treatment of the future implications of energy consumption implied
by the power rating of alternate bulbs. The utility coefficient on Cj thus determines the
degree to which the individual is time-consistent. More specifically, if the utility coefficient
on Cj implies that the consumer is willing-to-pay an amount equal to the present value of
savings associated with a 1W lower power rating, we term her time-consistent. We derive this
benchmark willingness-to-pay, W T PiR , for each consumer and a benchmark utility coefficient,
θiR , that would imply the individual would be willing to pay W T PiR . Note that when θi = θiR ,
the consumer is acting in a manner analogous to that when βu = 1 in the previous subsection.
Our calculation of W T PiR assumes an electricity price of $0.1147/kWh, as shared with
participants, and that a bulb is used 3 hours a day every day for eight years.28 Formally,
W T PiR =
T =8
X
Savingst
t=1
(1 + ri )t
(12)
In Equation 12, ri is the consumer’s temporal discount rate, and Savingst equals the annual
savings that accrue to the consumer upon choosing a bulb with a 1W lower power rating. In
general, we use an individual-level discount rate derived from an elicitation task embedded
in the experiment and designed by Kirby and Marakovic (1996). To test the sensitivity of
our results to the discount rate assumptions, we also use a common 6% discount rate.
26
It is especially helpful to temporarily assume away the direct effect of the Energy Star, since any given
behavioral instrument may not necessarily generate an analogous effect.
27
This is consistent with the treatment by Houde (2013), and we echo his assertion that this is in reality
open to debate. While we subsequently introduce it into our assumed decision utility function in Section
5.2, we do not discuss it here to focus instead on the impact of the Energy Star label on the coefficient on
energy consumption.
28
We implicitly assume that the rational consumer believes in expectation that electricity rates over these
eight years will remain constant.
20
W T PiR excludes any disutility from the environmental costs of consumption and should
be interpreted as a “selfishly” rational measure. We exclude these costs because the costs
incident on any one person should not make a first-order difference in W T PiR . We illustrate
this with an extreme example in which all electricity consumed is generated from coal-fired
facilities, which generally entail the largest external costs of electricity generation. In particular, Sundqvist (2004) suggests that the median external cost estimate in the literature is
$0.11/kWh, upon adjusting from 1998 to 2013 dollars. This reflects both localized costs, such
as those from mercury and particulate matter emissions, and distributed costs, such as those
associated with the emission of greenhouse gases. If we next assume that the $0.11/kWh
cost is exerted on only 100,000 people, any given person would include $0.000001/kWh into
his or her calculation of the value of energy consumption savings. While the total external
cost of $0.11/kWh should be considered in the evaluation of social welfare, the derivation
of a given individual’s “selfishly” rational assessment will be essentially unchanged upon
including his or her share of the environmental costs.29 We incorporate the full external cost
in Section 5 where we illustrate a comprehensive valuation of the Energy Star program.
The estimated marginal utility of wealth, ηi , points to a benchmark utility coefficient,
θiR , that would imply the individual would be willing to pay W T PiR :
θiR = W T PiR · ηi
(13)
We define the wedge, wedgei , as the difference between θiR and θi , the estimated coefficient
on energy consumption in the absence of the Energy Star:
wedgei = θiR − θi
(14)
With θiR , we can update our expression for experienced utility to:
Ueij = ηi · pj + θiR · Cj + ΥT Xj
(15)
Equation 15 allows us to determine the utility experienced by the time-consistent consumer
with her utility-maximizing choice, x∗ . This level of utility serves as the benchmark against
which we compare the utility experienced by time-inconsistent decision-makers. When Equa29
Of course, the distribution of cost incidence may not be uniform, as we have assumed, but a very
drastically different distribution would be required for the implication here to change.
21
tion 11 has θi 6= θiR , it should be interpreted as the decision utility function. In the presence
of the Energy Star, the decision utility function includes an additional term, γi · Cj .
The consumer who acts as if she were applying time-inconsistent preferences would display choice behavior consistent with the maximization of a utility function with a coefficient
θi 6= θiR . The wedgei term allows us to quantify the degree to which the consumer is timeinconsistent and identify the nature of inconsistency; in particular, the consumer can be
either present-biased (θi < θiR ) and under-value savings on energy consumption or futurebiased (θi > θiR ) and over-value them. Continuing our assumption that the consumer faces
a continuous choice set and thus changes her choices with any change in wedge value, as
|wedgei | approaches zero, the choice of the time-inconsistent individual, xd , becomes more
similar to x∗ . Consequently, the internality or loss in experienced utility, Ue (x∗ ) − Ue (xd ),
will become smaller. Thus, to evaluate whether the Energy Star provides value or harms a
particular individual, we determine whether or not its effect is to shift her marginal utility
coefficient closer to the rational benchmark.30
Though we constrain the utility coefficients on energy consumption to be negative, for
ease of interpretation we discuss the negative of the coefficients. Thus, individuals with
positive wedges undervalue savings on energy consumption, and those with negative wedges
overvalue savings on energy consumption.31 We discuss γi similarly; the higher it is, the
greater is the impact of the Energy Star in increasing an individual’s valuation of savings on
energy consumption.
To gain an intuition for the effect of the Energy Star, we define the mitigation, miti ,
performed by the Energy Star logo as a link between the effect of the Energy Star on wedgei
and implications for experienced utility. We measure mitigation as the ratio of the coefficient
30
Note that our assumption of a continuous choice set in which choices change as wedges change is critical
in our exposition here. To evaluate the Energy Star program more generally, we need to know if its impact
on decision utility coefficients actually affects choices made and thus the experienced utility obtained by the
consumer. We drop the simplifying assumption and present a more nuanced discussion of the effects of the
Energy Star on experienced utility in Section 5.2.
31
In the context of appliance choice, a natural retort to the concept of overvaluation of energy savings is
that a large coefficient on energy consumption is also consistent with strong preferences for environmental
protection. While that is true, even those individuals who have strong environmental preferences may
overvalue savings on energy consumption. Since environmental quality can be affected by many actions,
the individual can still pursue too much environmental conservation in appliance choice relative to other
opportunities. Stated otherwise, the individual may have lower cost ways by which to achieve the same level
of environmental conservation.
22
on the Energy Star and Consumption interaction term, γi , to wedgei :
miti =
γi
γi
= R
wedgei
(θi − θi )
(16)
Note that the mitigation measure is only defined when a wedge exists; it is undefined when
θiR = θi , since by definition no wedge exists to be mitigated. When an individual has
wedgei = 0 in the absence of a behavioral instrument like the Energy Star, any impact of
the instrument on the apparent utility coefficients characterizing his choices implies a loss in
experienced utility away from the normative utility level.32
Figure 3: The mitigation scale is split into three areas: (1) wedge exacerbation, (2) wedge
mitigation, and (3) an overshoot in mitigation. The first and second areas imply only
losses and gains in experienced utility, respectively. In the third area, mitigation increases
experienced utility in the range 1 to 2 but decreases it otherwise. The threshold value of 2
reflects a symmetric choice set.
Four features illustrate the connection between the impact of the Energy Star on wedges
and experienced utility. First, when miti < 0, the Energy Star changes the utility coefficient
on energy consumption in the opposite direction of that needed to set wedgei = 0. This
implies that the Energy Star yields a loss in experienced utility. The Energy Star delivers the
largest gain in consumer utility when it exactly closes the wedge, or when miti = 1. When
0 < miti < 1, the Energy Star partially closes the wedge and increases consumer utility.
When miti > 1, the Energy Star overcompensates for the wedge; the sign of the wedge flips,
with a wedge between θiR and θi + γi remaining. As long as the new wedge does not imply
a larger loss in experienced utility than implied by the old wedge, the Energy Star can still
improve the consumer’s utility. The threshold mitigation value will depend on the nature of
32
This follows from our assumption of a continuous choice set. Changes in the utility coefficients that do
not imply changes in the alternative selected will not generate these losses.
23
the choice set, since that will drive the implications for experienced utility. Assuming that
the choice set is symmetric, in the sense that deviations of equal magnitude from θiR in either
direction imply equal losses in experienced utility, the Energy Star will improve experienced
utility as long as 1 < miti < 2. If miti > 2, the overcompensation is sufficiently severe that
the resultant wedge in the opposite direction implies a loss in experienced utility.33
Figures 4 and 5 plot the wedges and mitigation levels observed in our sample. Analyzed
in isolation, the means in both plots corroborate the motivation for the Energy Star program
and imply that it accomplishes its goal of shifting individuals’ utility coefficients on energy
consumption toward the “rational” level.34 On average, participants’ utility coefficients
on energy consumption are smaller than their rational benchmarks, and mitigation by the
Energy Star is consistent with improvements in experienced utility. These average trends
are in line with the estimates shared in Section 3.2.
Figure 4: The observed wedges imply substantial heterogeneity in departures from the rational benchmark. Though the mean and median wedges equal 0.47 and 0.09, respectively,
a substantial portion (37%) of the wedges are negative.
The distributions of wedge and mitigation values around the average levels indicate that
33
If the choice set available to the consumer implies an asymmetry around θiR , the analyst should re-
calibrate this cutoff value.
34
Figure 5 is truncated at ±5. The mitigation levels range from -66.4 to 86.9.
24
Figure 5: The observed mitigation levels imply substantial heterogeneity in consumers’ response to the Energy Star label. The plot is truncated at ±5, though the observed mitigation
ranges from -66.4 to 86.9.
there are important differences in individual-level responses to the Energy Star. Our framework in Figure 6 categorizes the utility ramifications of the heterogeneous impacts of the
Energy Star. The framework groups participants into one of four groups defined by the
signs of wedgei and γi . These groups reflect the four possible effects of the Energy Star:
people either under- or over-value savings on energy consumption, and the Energy Star either increases or decreases this tendency. Group membership reflects the experienced utility
ramifications of these effects. Individuals in Groups A and D are those who, in the absence
of the Energy Star, over-value savings on energy consumption; formally, θi > θiR . From the
perspective of wedge mitigation, the desired impact of the Energy Star would be to decrease
θi closer to θiR . Consequently, those for whom γi < 0 could realize a gain in experienced
utility gain, while those for whom γi > 0 are subject to a loss in experienced utility. The
tentative language for the former group reflects the possibility that γi could be sufficiently
negative that it implies miti > 2 for the individual. As Figure 3 shows, this would lead to
a loss in experienced utility. In contrast, participants in Groups B and C are those who,
in the absence of the Energy Star, under-value savings on energy consumption, such that
25
Figure 6: Participants are binned into four groups defined by wedgei and γi . The groups in
red font are those for which the Energy Star implies a decrease in individual utility, since
miti < 0. The numbers below the group labels indicate the number of participants in our
sample of 1,550 binned into the group.
θi > θiR . For these individuals, the “desired” impact of the Energy Star is that γi > 0. Again,
if γi is too much greater than 0, individuals in Group C could bear a loss in experienced
utility. However, those in Group B unambiguously face a decrease in experienced utility, as
the impact of the Energy Star is in the direction opposite of that needed to close the wedge.
Note that the diagonal including Groups A and C includes those for whom miti is greater
than 0; for these individuals, the Energy Star could increase utility. The other diagonal with
Groups B and D includes individuals for whom miti is smaller than 0 and who experience a
loss in utility.
In our sample, 60, 146, 829, and 515 participants are in Groups A, B, C, and D, respec26
tively.35 We highlight two points. First, we observe that more individuals have miti > 0
than the opposite sign. In particular, the Energy Star affects the utility coefficient on consumption in the desired direction among 57% of the sample, or 889 individuals. These counts
would suggest an overall improvement in consumer welfare, but the actual implications will
depend on the marginal utilities of wealth. Second, the Energy Star acts to increase the
valuation of energy consumption savings for 87% of individuals. This is consistent with the
mean coefficient on the Energy Star and Consumption interaction term in Table 1.
We summarize in Finding 2 the implications of the Energy Star for experienced utility.
Finding 2: Assuming a continuous light bulb choice set, that choices are known with certainty, and that consumers’ decision utility functions do not include the “brand” effect of
the Energy Star, the impact of the Energy Star is to exacerbate wedges among and decrease
experienced utility for 661 (43%) of the 1550 individuals in our sample. It increases experienced utility by affecting wedges in the desired direction for at most 889 (57%) of the 1550
individuals in our sample.
The overall value of the program will depend on the extent of mitigation among individuals, their marginal utilities of wealth, and the impacts the Energy Star has on experienced
utility and environmental externality mitigation. We next turn to those aspects.
5
The Value of the Energy Star to Consumers
5.1
Evaluation Methodology
Our discussion in the previous section highlighted that the Energy Star label can influence
a change in the good that the individual purchases. Since the alternative selected in the
presence of the Energy Star program may imply experienced utility and external costs of
electricity generation that are higher or lower than those obtained in its absence, we must
value both changes. Equation 17 captures the two effects:36
35
These counts are based on rational benchmarks defined by individual-level discount rates. If we use a
common 6% discount rate, the group counts are 41, 165, 879, and 475, respectively.
36
While the equation will appear exactly the same as one that traditional public economics would prescribe,
the consumer surplus terms reflect experienced utility levels stemming from different decision utility functions
in the absence and presence of the instrument. In contrast, standard public economics would assume that
surplus in both settings would reflect a stable experienced utility function.
27
exp
exp
V alueES = (CSES
− CSN
ES ) − (ECEES − ECEN ES )
(17)
In the expression above, CS exp represents the consumer surplus based on the experienced utility from bulb purchase, ES the setting in which the Energy Star program exists, N ES that
in which it does not, and ECE the external costs of electricity generation and consumption.
Changes in ECE enter negatively because consumers benefit from smaller external costs.
Equation 17 reflects several assumptions. Our use of a traditional public economics valuation approach assumes that consumers incorrectly consider only the energy consumption
attribute, such that the utility coefficients on the other attributes can be used to approximate
the consumer’s experienced utility. Moreover, notwithstanding the findings by Houde (2013)
that firms respond strategically to the Energy Star certification, we assume a competitive
supply side that sells the same products with and without the Energy Star. We further
assume that firms do not modify the prices of their products in response to consumers’
valuation of the Energy Star label itself.
To determine the change in the consumer surplus, we evaluate the expression below:
∆CS exp =
X
i
−∆
Internalityi X 1 ES
=
(Ueij∗ − UeNij∗ES )
ηi
η
i
i
(18)
The division by ηi allows us to derive a dollar-denominated metric from the util-denominated
measure of experienced utility, Ue . The j∗ index denotes the good selected by the consumer.37
To calculate the change in the external cost of electricity consumption, we must determine
the external cost of each kilowatt hour of electricity consumed. We assign each product
an external environmental cost, ECEj . We derive this by first calculating the electricity
consumed by a 1W bulb operating for 3 hours a day for 8 years and subsequently scale
this figure by both the actual power rating of bulb j and ED, the environmental damage
per kWh of electricity consumption. We discount future external costs using a 6% social
discount rate, rs .38 Formally:
37
The negative sign preceding the ∆ Internality term follows from the definition of the internality as the
loss in experienced utility.
38
Since the marginal costs of greenhouse gas emissions reflect the aggregate stock of emissions, they increase
at the rate of inflation. Strictly speaking, we should not discount these costs by the social discount rate.
However, we use an aggregated measure of external costs; though a disaggregation would refine our valuation
estimates, it does not add to the central point of the paper.
28
ECEj =
T =8
X
Cj (3∗365)
ED
1000
(1 + rs )t
t=1
(19)
We assume that ED equals $0.068/kWh. This reflects a weighted average of damage estimates for electricity derived from coal, oil, natural gas, nuclear, hydro, wind, and solar
sources, with weights determined by the share of U.S. electricity generation contributed by
each source.39
The evaluation methodology we have thus far outlined is based on the intuition developed
in Section 4.2 for how the Energy Star would work under assumptions that we know the
consumers’ choices with certainty, that the consumer accesses a continuous light bulb choice
set, and that consumers’ decision utility functions do not include the direct “brand” effect of
the Energy Star label. In the following subsection, we drop these assumptions and discuss the
necessary modifications to our evaluation of the Energy Star program. The final subsection
offers an evaluation for a particular choice set.
5.2
Three Complexities
Our evaluation methodology requires us to simulate consumer choice in both the absence
and presence of the Energy Star program. As Figures 7 and 8 illustrate, the extent of
internalities depends on the choice set available to the consumer and on the curvature of the
experienced utility curve when plotted against the value of the utility coefficient, θi .40 Figure
7 depicts a situation in which there are large potential losses from internalities. Assume that
the consumer acts as if she applies a decision utility coefficient of θi0 in the absence of the
Energy Star. ∆Ue1 describes the increase in experienced utility associated with an increase
in the coefficient by the quantity γi1 and towards θiR . As the figure shows, the Energy Star
39
We scale the estimates from 1998 to 2013 dollars using the U.S. BLS Inflation Calculator. For each
source, we use the median level of damages reviewed by Sundqvist (2004) and share of electricity generation
in the U.S. from the U.S. Energy Information Agency (2013). In particular, we use data from a table of Net
Generation by Energy Source. We allot generation from “petroleum liquids” to oil and remove the small
amount of generation from petcoke, “other gas” and biomass. The median figures mask many underlying
differences across studies, such as the inclusion or exclusion of CO2 prices, the consideration of full fuel cycle
effects, estimations by either abatement cost or damage cost approaches, and top-down versus bottom-up
estimates (Sundqvist, 2004). However, the cost is in the range of others used; for example, Houde (2013)
uses $0.079/kWh, which reflects a $67.1/t damage from CO2 emissions.
40
The appendix includes a discussion that links the utility coefficient to the choice of products and ultimately to the experienced utility curve.
29
label can also imply losses in experienced utility. ∆Ue2 denotes the lost experienced utility
associated with an increase in the coefficient by the quantity γi2 and away from θiR . As the
consumer acts as if she applies a θi that is farther away from her θiR in her decision making,
she chooses goods that imply a larger loss in experienced utility. This situation corresponds
to one in which one light bulb does not dominate the others on the basis of both price and
consumption.
Figure 8 depicts a scenario in which one alternative in the choice set dominates the others;
namely, one bulb dominates the others on both price and consumption. Regardless of the
θi that the consumer applies, her experienced utility remains the same, and the internality
cost, zero. Thus, ∆Ue1 = ∆Ue2 = 0.
Figure 7: Implications for experienced utility when the choice set available to the consumer
implies a highly non-linear experienced utility versus θi curve. This scenario arises when
there is a rich trade-off between up-front prices and operating costs.
The difference between Figures 7 and 8 underscores a fundamental point: the greater the
curvature in the experienced utility versus θi curve, the greater the loss from mis-optimization
by consumers and the greater the potential gain from a ‘nudge’ by the Energy Star program
30
Figure 8: Implications for experienced utility when the choice set available to the consumer
implies a horizontal experienced utility versus θi curve. This scenario arises when the choice
set includes one alternative that dominates the others.
of θi in the right direction. The sensitivity of the implications of a wedge between θi and θiR
implies that the analyst can provide only an evaluation of the Energy Star program that is
conditional on the assumed choice set. Since the consumer may evaluate alternatives unseen
and unconsidered by the analyst, the latter is unable to observe the former’s “true” Ueij (θi )
curve. Of course, this possibility exists even when an analyst uses data on alternatives
available in-market. In this case, the weak concavity of the experienced utility curve implies
that estimates of the loss attributable to the Energy Star program can be interpreted as
lower bounds.41
The logic connecting the nature of the choice set to implications for the change in experienced utility extends to implications for externalities as well. If the choice set includes one
dominant bulb, the impact of the Energy Star program on θi will not yield any changes in the
41
We cannot interpret estimates of the gain as upper bounds because it is possible that alternatives yielding
a higher experienced utility exist within the consumer’s choice set but were unobserved by the analyst.
31
alternative actually selected and, consequently, on the extent of externalities. In the setting
in which the bulb selected is sensitive to the utility coefficient, however, we can identify a
useful inequality to approximate whether the effect of the Energy Star on a given consumer
implies a social gain or a social loss. When the effect of the Energy Star program is to
increase θi closer to θiR such that the consumer is more likely to purchase a bulb of a lower
power rating, there is an unambiguous weak gain in social welfare, since the consumer’s
experienced utility weakly increases and the external costs of electricity consumption weakly
decrease. Similarly, when the program decreases θi away from θiR , there is an unambiguous
weak loss in social welfare, as the consumer’s experienced utility weakly decreases and the
external costs of electricity consumption weakly increase. There are two situations in which
there is some ambiguity about the social implications of the Energy Star program. These
arise when the program either decreases θi toward the rational benchmark and thus encourages more consumption or increases θi away from the rational benchmark. In this case, we
must compare the gain or loss in experienced utility with the gain or loss in environmental
externalities. With “small” changes in the utility coefficient, the difference in experienced
utility can be approximated by
(θi +γi )
ηi
ES
· Cj∗
(θi + γi ) −
θi
ηi
N ES
· Cj∗
(θi ) while the change in
ES
N ES
externality is given by ECEj∗
(θi + γi ) − ECEj∗
(θi ). For small changes in the utility
coefficient, the analyst can thus compare the approximated difference in experienced utility,
which we have expressed in surplus terms to keep units consistent, to the difference in the
external costs of electricity consumption.
A second complexity arises because we do not know with certainty the choices that the
consumer will make. We must assume a deterministic setting for Equation 17 to hold.
However, our utility parameters were derived only with an assumption about the nature of
unobserved utility. Consequently, and as Small and Rosen (1981) and Leard (2013) highlight,
Equation 17 must be evaluated in expectation. We thus re-express that equation as follows:
exp
exp
E[V alueES ] = E[(CSES
− CSN
ES )] − E[(ECEES − ECEN ES )]
To evaluate the above expression, we evaluate both Equation 20, which computes the
expected change in the external costs of electricity generation, and Equation 21:
E[∆ECE] =
XX
(PijES − PijN ES ) · ECEj
i
j
32
(20)
E[∆CS exp ] =
XX
i
(PijES − PijN ES ) ·
j
Ueij
ηi
(21)
In the equations above, PijES and PijN ES are the simulated probabilities that participant i
selects alternative j when the Energy Star program exists and when it does not, respectively.
This addition of the probability terms implies that the Energy Star imposes costs and benefits
in expectation. Said otherwise, it can change the probability with which the consumer
chooses a given alternative in her choice set. For example, if a consumer’s optimal choice
entails the lowest consumption but highest priced good available in a choice set and the
Energy Star acts to decrease the probability with which she selects this good by reducing
the magnitude of the coefficient on energy consumption in her decision utility function, it
exerts a cost in expectation.
The third addition to our evaluation analysis is the effect of the Energy Star on decision
utility, independent of its effects on θi . This implies an update to our decision utility function
in the presence of the Energy Star:
Udij = ηi · pj + (θi + γi ) · Cj + ΥT Xj + λi · ESj
(22)
The impact of this consumption-independent effect of the Energy Star is ambiguous and
must be empirically evaluated for a specific choice set. Since only 60 of the 1550 individuals
in our sample had a value of λi < 0, we can characterize the modal impact of this Energy
Star effect. If the utility-maximizing alternative is Energy Star certified, this Energy Star
effect will simply act to increase the probability with which the optimal good is selected. If,
however, the utility-maximizing alternative is not certified, this effect will have the opposite
implications and imply a lower expected experienced utility. To the extent that this Energy
Star effect compels individuals to select goods at least as energy efficient as they would
have in its absence, it contributes to a weak reduction in environmental externalities. In the
subsequent subsection, we quantify the incremental impact of the consumption-independent
effect of the Energy Star on consumer welfare.
5.3
The Value of the Energy Star
To evaluate the expected change in value, we apply the individual-level utility coefficients to
simulate consumer choice in the absence and presence of the Energy Star. The simulation
33
yields choice probabilities for both contexts. We use a hypothetical choice set containing
CFLs available for purchase from a large national retailer in early 2014.42 The bulbs have a
narrow band of power ratings (13 – 16W), since we restricted the lumen range of the choice
set to remain between 800 and 900 lumens.
We perform three versions of the evaluation procedure. In the first version, we isolate the
impact of the Energy Star on consumers’ utility coefficients on energy consumption. In the
second, we include the energy-independent effect of the Energy Star in consumers’ decision
utility. The final version restricts the effect of the Energy Star on the utility coefficient
on energy consumption to bulbs that are Energy Star certified. This final version is most
consistent with our econometric estimation. The other versions require an assumption that
in an actual market, the presence of the Energy Star label prompts changes in consumers’
apparent decision utility functions across both labeled and unlabeled goods.
Table 2 summarizes the value of the Energy Star for the three versions described above.
The top block of rows details the expected gain in experienced utility, and the middle block,
the expected reduction in external cost. The “Total Increase in Value” is the sum of the
two “All groups” totals listed below in each block of rows. At the bottom of the table,
we scale the value of the Energy Star to our participants to the entire population of U.S.
households. This row estimates the value per year across owner-occupied households in the
U.S., assuming they would respond as the participants of our experiment did and purchase
four bulbs each in a given year.43
We make three comments based on Table 2. First, the increases in experienced utility
do not match the theoretical guidance provided by Figure 6. In particular, we would have
expected utility losses among those in Groups B and D and gains only among those in
Groups A and C. The discrepancy between expected and actual results stems from the fact
that we are evaluating changes in expectation. In the choice set we use for the simulation,
one alternative dominates the others by offering the lowest consumption and lowest price
option. This good is the experienced utility maximizing alternative for all individuals in
42
In particular, we derived this choice set from the larger choice set available to consumers at
www.lowes.com, the online storefront for the retailer. We exclude certain bulbs with warmth levels beyond those considered in the choice experiment, since we are unable to estimate the marginal utility of this
level. We also exclude decorative bulbs, or those whose light is a color other than white.
43
We scale the value of the instrument by the product of (1) the ratio of the total number of U.S households
(approximately 115 million) to the 1550 households in our sample and (2) the share of households that are
owner-occupied. We use 65.2% for the latter, per data from the U.S. Census Bureau (2014).
34
Version 3:
Version 1:
Version 2:
Indirect ES
Both ES
effect only
effects
Group A: ES reduces positive wedges
-$2
$16
$25
Group B: ES increases negative wedges
-$8
$23
$54
Group C: ES reduces negative wedges
$11
$96
$42
Group D: ES increases positive wedges
$13
$178
$69
$13
$313
$189
$0.01
$0.20
$0.12
Group A
-$1
$1
$3
Group B
-$3
$1
$6
Group C
$3
$12
$6
Group D
$4
$16
$11
$4
$30
$25
$0.00
$0.02
$0.02
$17
$343
$215
$0.01
$0.22
$0.14
$4M
$60M
$36M
Both ES effects;
Indirect effect on
labeled bulbs only
Exp. Gain in Experienced Utility
All groups
Exp. Reduction in External Cost
All groups
Total Increase in Value
Scaled to all U.S. Households
Assumes four bulbs purchased per year
Table 2: In the “Total” rows, the top number quantifies the impact across the entire population and the bottom number, the per-capita impact. The final row scales the impact across
the 1,550 households in the study to the population of U.S. owner-occupied households.
Differences between totals and the sum of group-specific amounts reflect rounding errors.
Note: “Exp.” denotes “Expected,” “ES,” “Energy Star,” and “M,” million. Recall that the
indirect effect of the Energy Star is that on the utility coefficient on energy consumption
and that the direct effect is the “brand” effect that is independent of energy implications.
our sample.44 The increase in value observed in all three versions reflects the fact that the
44
While is may be surprising that a commercially available choice set would include both a dominant bulb
and others, we note that the dominance follows upon considering only price, consumption, and bulb warmth.
Our specifications of the utility functions abstract away from brand considerations that may be relevant to
the consumer. Moreover, our example evaluation retains the assumption of our stated choice experiment
that the lifetime of all bulbs is equal. In the choice set used for simulation, the bulbs’ lifetime ranged from
7 to 11 years. Thus, the bulb we consider dominant may not be, given a more comprehensive description of
35
Energy Star program more often increases the probability with which consumers select this
experienced-utility maximizing good than it decreases it. Consumers in Groups A and B,
who act as if their decision utility coefficient on energy consumption decreases in the presence
of the Energy Star, experience a loss in expected experienced utility in Version 1 because the
probability with which they select this alternative decreases. In contrast, those in Groups
C and D act as if their decision utility coefficient on the attribute has increased, and the
relevant probability increases.
A comparison of Version 2 with Version 1 reveals the powerful impact of the energyindependent effect of the Energy Star label. Since the dominant good in the choice set is
labeled with the Energy Star logo, we observe that individuals in all four groups experience
a larger gain in experienced utility in Version 2 than in Version 1. The size of the energyindependent Energy Star effect is sufficiently large that its impact on the probabilities of
selecting the alternatives more than compensates for the decrease in probability of selecting
the experienced utility maximizing good among Groups A and B in Version 1. Version
3 retains the assumption that the Energy Star exerts this effect but limits changes in the
decision utility coefficient on energy consumption to those bulbs that are Energy Star labeled.
Though the overall effect of this restriction is a reduction in the total increase in value in
Version 3 relative to Version 2, those in Groups A and B obtain an increase in their expected
gain in experienced utility and those in Groups C and D, a decrease. In Version 3 and for
those in Groups A and B, the impact of the Energy Star on the energy consumption utility
coefficient is equivalent to imparting a decrease in the perceived consumption level of labeled
goods, relative to those without the label. The Energy Star exerts the opposite effect on
individuals in Groups C and D.
We note finally that the expected impact on experienced utility is larger than that on
external costs. Though instruments such as the Energy Star may be motivated by their
potential to reduce external costs, their impact on private costs may be bigger. On the one
hand, this reflects our decision to restrict the simulation bulb choice set to bulbs of 800 –
1000 lumens. On the other hand, this restriction may reflect how consumers make their bulb
purchase decisions: bulbs are installed in a variety of locations throughout the house, and
the search for bulbs may be limited to those within a particular lumen range and therefore
a specific range of brightness. Since the variation in power requirements for a given range is
individuals’ apparent utility functions.
36
relatively small, the reductions in external cost are unlikely to be large. Finally, though we
observe an expected increase in value from the Energy Star, conditional on the simulation
choice set we used, our numerical example does not imply that the Energy Star will perform
this way with any choice set. An example choice set in which we would calculate a negative
value is that with all bulbs of consumption between 5W and 16W in 0.25W increments
associated with all prices from $11.50 to $0.50, respectively, in -$0.25 increments.
The Energy Star appears to be a viable instrument if the goal of policymakers is solely
to reduce externalities. Nonetheless, the instrument has the potential to exert large costs
on consumers that are attributable to changes in their experienced utility. More broadly,
in contexts where societal goals of mitigating market failures could be in conflict with individuals’ experienced utility, the analyst should consider both effects in determining the
attractiveness of a behavioral instrument. Finding 3 summarizes our findings:
Finding 3: Using a choice set in which one alternative dominates the others, we find that
the Energy Star improves consumer welfare in expectation. This expected improvement stems
more from the impacts of the instrument on experienced utility than those on external costs.
We emphasize that the unconditional value of the Energy Star is ambiguous.
6
Conclusion
Given patterns of under-investment in energy efficient goods, the analyst with a goal of
reducing environmental externalities could conclude that the more often the Energy Star
implies an increase in the valuation of savings on energy consumption, the more valuable
the program. However, this would ignore the effect of the Energy Star on individual welfare.
For any given individual, the effect of the Energy Star is ambiguous. At one level, it is
unclear if the Energy Star will actually increase the attention paid to energy consumption
information, since the label could instead provide a license to ignore the attribute. This
implies a second level of indeterminacy: the impact of the Energy Star may be to prompt
the individual to make choices that reflect a lower or higher valuation of savings on energy
consumption, relative to that expressed in the absence of the label.
Ideally, the Energy Star program would not only address environmental externalities
but also encourage consumers to make choices that strike a balance between the upfront
price and long-run costs of alternate goods that is more aligned with utility maximization.
37
To do so, it would address tendencies among consumers to under- or over-value savings on
energy consumption. Such over- or under-valuation could exert internalities on consumers
if it compels them to select goods that imply changes in their experienced utility. The
reduction of these internalities would increase individual welfare. Since an under- or overvaluation of energy consumption savings can exert internalities on the consumer, any impact
of the Energy Star on the valuation of energy consumption savings and expected changes in
experienced utility must be quantified in policy analysis.
This realization implies the question that motivated this paper: given its effects on both
externalities and internalites, what is the value of the Energy Star to consumers? This
paper presented and analyzed data from a stated choice experiment intended to answer this
question. Our chief contribution was an analytic framework that uses these data to quantify
the impact of the Energy Star on internalities and to value the program. The fundamental
strategy in our framework is to establish a benchmark for experienced utility coefficients that
reflect standard time-consistent preferences. Specifically, we used our data to estimate utility
coefficients in the absence and presence of the Energy Star. Assuming a particular price
of electricity and lifetime of the bulb products considered by our participants, we derived
a rational benchmark for the utility coefficient on energy consumption. The benchmark
coefficient is analogous to βu = 1 in the beta-delta conceptualization presented in the paper,
since it is the parameter consistent with the maximization of an experienced utility function
by the consumer. We quantified the impact of the Energy Star on internality losses by
comparing the difference in expected experienced utility in the absence and presence of
the Energy Star, relative to that predicted with the rational benchmark. Our discussion
emphasized that wedges between the rational benchmark coefficient on energy consumption
and those observed in the presence and absence of the Energy Star are not themselves
sufficient to imply a loss to the consumer. Such losses occur only if the wedge compels
the consumer to select a good other than the one that maximizes experienced utility. The
same commentary applies to changes in externalities. Given this, our methodology can value
internality and externality effects, conditional on a given choice set.
This framework can be used in other contexts upon making the two assumptions we
have made. Namely, the analyst must first assume that the utility coefficients on the attributes other than those affected by the behavioral instrument reflect standard preferences.
The second assumption establishes a rational benchmark experienced utility against which
38
internality losses in the absence and presence of the Energy Star can be measured. The
framework will be more difficult to apply in contexts in which it is not as straightforward to
measure the experienced utility implications of choosing a given alternative.
Our results suggest that the impact of the Energy Star on internalities is larger than that
on environmental externalities. We applied our evaluation methodology to a choice set of
light bulbs offered by a national retailer. When evaluating expected changes in internalities
and externalities, we find that the Energy Star delivers modest reductions in the expected
cost of both. This result does not generalize to all choice sets subject to the influence of
the Energy Star. If the choice set entails a richer trade-off between up-front prices and
costs of operation, a poor performance of the Energy Star would stem from the situations in
which it increases and decreases the magnitude of consumers’ utility coefficients on energy
consumption among those who, in its absence, already overvalue or undervalue consumption
savings, respectively. For these individuals, the Energy Star would significantly increase
internality losses when selecting from such a choice set.
The potential for non-pecuniary measures such as the Energy Star to impose internality
costs implies that such instruments can yield negative dividends. That is, the instruments
may deliver greater internality losses than externality gains. This adds a nuance to the
assertion by (Allcott, Mullainathan, and Taubinsky, 2014) that taxes aimed at the reduction
of externalities could yield a “double dividend” by also addressing internalities. Non-price
instruments are receiving increasing attention in light of popular discussions of strategies to
“nudge” consumers. Our finding here should encourage a fuller evaluation of such proposals
with a careful accounting of their impacts on internalities and consumer welfare.
39
7
Appendix
Distribution of individual-level coefficients on the ML interaction term
Figure 9: The mean individual-level utility coefficient on the Energy Star and energy consumption interaction term, -0.126, implies that the Energy Star increases the weight consumers place on energy consumption. However, the distribution reveals that the label both
increases and decreases this weight. The label decreases the weight for the 13% of individuals
with an estimated coefficient on the interaction term greater than zero.
40
Interpretive information provided in the stated choice experiment
Description of the task
We are interested in understanding how you purchase light bulbs, and we will ask you a
series of questions in which you will pick your preferred bulb among three alternate light
bulbs.
As you answer these questions, please assume that all light bulbs:
1. Are compatible with standard fixtures (for example, ceiling or lamp fixtures)
2. Are dimmable; that is, the intensity of their output can be adjusted by using, for
example, a light switch that allows one to choose a level of light intensity
3. Have an equal delay between when you turn them on and when they reach full brightness
4. Have an equal brightness
5. Have the same expected lifetime. All bulbs in this study are compact fluorescent light
bulbs (CFLs). CFLs last approximately 8,000 hours, while standard incandescent bulbs
typically last 750 hours
The light bulbs we will ask you to compare will differ along some or all of the following
dimensions:
Price The price of the light bulb is the price you would pay in a retail store.
Light color The light color is a measure of the color temperature of the light bulb. In
residential settings, most light bulbs produce warm white, soft white, neutral or cool light.
The color temperature profiles of the light bulbs we will describe shortly are described below.
Energy Star Certification Energy Star certified light bulbs use about 75% less energy than
traditional incandescent bulbs and last at least 6 times longer. The Energy Star program is
a voluntary certification program, and manufacturers may elect to apply labels such as the
one below to products that meet energy consumption criteria set by the U.S. Environmental
Protection Agency (EPA). Since the program is voluntary, some light bulbs that meet the
criteria do not carry the label. An example label appears below.
We will use the following label for light bulbs that are not Energy Star labeled; remember
that a bulb may be unlabeled either because (1) it does not meet the federal standard for
41
Energy Star certification or (2) although it meets the federal standard for Energy Star
certification, the manufacturer has opted to not certify it.
[The instructions displayed the Energy Star label and a blank blue label used in the
experiment to denote a non-certified bulb. Below, we include the three versions of details
provided to participants in experimental conditions 1, 2, and 3, respectively.]
Consumption: Energy Demand (W) The energy demand refers to the ‘watt’ rating of the
bulb. Watts are a standard unit of power and measure the rate at which a light bulb
consumes energy. For a given amount of light outputted, different light bulb technologies
are characterized by different watt ratings.
Example: a low watt CFL light bulb can produce an equivalent light intensity as a high
watt traditional incandescent light bulb.
Consumption: Energy Demand (Estimated Annual Cost) The estimated annual energy cost
refers to the monetary cost of energy consumed by a light bulb in one year. The actual
energy cost for a bulb will vary depending on how often it is used and the price of electricity.
Information is provided about the annual cost of electricity associated with a given bulb.
To derive this annual cost, we assume that the light bulb is used 3 hours a day and use an
average U.S. residential electricity rate of $0.1147/kWh.
Consumption: Energy Demand (Watts and Estimated Annual Cost)
The energy demand refers to the ‘watt’ rating of the bulb. Watts are a standard unit of
power and measure the rate at which a light bulb consumes energy. For a given amount of
light outputted, different light bulb technologies are characterized by different watt ratings.
Example: a low watt CFL light bulb can produce an equivalent light intensity as a high
watt traditional incandescent light bulb.
Information is also provided about the estimated annual cost of electricity associated
with a given bulb. To derive this annual cost, we assume that the light bulb is used 3 hours
a day and use an average U.S. residential electricity rate of $0.1147/kWh.
Details of the scale analysis
Our application of the Swait and Louviere analysis begins by testing whether one can
reject the hypothesis of equal scale factors across the three experimental conditions. Though
the scale factor cannot be identified, the ratio of the scale factor within the data from
one experimental condition to that within another can be identified. We extend the Swait
42
and Louviere hypothesis testing procedure to three experimental conditions, wherein the
overall hypothesis is one of coefficient and scale parameter equality: H1 : β1 = β2 = β3
and σ1 = σ2 = σ3 . To test the hypothesis, we first estimate separate parameter vectors β3 ,
σ2 · β2 , σ1 · β1 , which provide log likelihoods of L1 , L2 , and L3 , respectively.45 In a second
step, we impose a constraint that β1 = β2 = β3 but allow σ1 and σ2 to vary.46 This second
estimation generates a single log likelihood estimate, Lstep2 , and we form a test statistic
λ1 = −2 [Lstep2 − (L1 + L2 + L3 )], which is asymptotically distributed with K + 2 degrees of
freedom, where K is equal to 7. Table 3 summarizes our log likelihood estimates and informs
our rejection of H1.
L3
-8996.128
L1
-8909.386
L2
-8850.386
LStep2
-26850.037
λ1
p-value
σ1
σ2
188.270
<0.001
0.960
0.935
Table 3: Summary output from the test of coefficient and scale parameter equality
Given the rejection of H1, we follow the suggestion by Swait and Louviere to scale our
data by the σ1 and σ2 values that yielded the maximum log likelihood in the step 2 estimate.
Table 3 details these scale factors in the last two columns.
An example to illustrate the origin of internalities
Consider an infrequent caffeine consumer who benefits from an immediate large gain in
mental acuity when she consumes caffeine but suffers, with delay, from considerable caffeineinduced anxiety. Facing a work deadline that has limited her sleep recently, she is deciding
whether to consume an espresso drink with 1 or 2 shots of espresso. The outside option,
which is associated with a baseline utility of zero utils, is to skip the drink altogether;
we refer to this as the zero shot option. Table 4 presents the net benefit, in units of utility
experienced, relevant to her decision. The consumer instantaneously experiences the benefits
of increased alertness in time one and the disutility of the caffeine-induced anxiety at time
two. The implication is that the latter costs are delayed while the benefit of consumption
is immediately realized. This allows the costs to be only partially salient to the consumer
making a choice at time one. Recall that the time-zero decision-maker, in contrast, is assumed
45
Condition 3 serves as our reference condition because it includes both descriptions of consumption
information.
46
We search in a grid from 0.01 to 2.25 with a step size of 0.025.
43
to fully consider both the future benefits and costs of the alternatives available to her.
Shots of espresso
Time = 1
Time = 2
Zero
0
0
One
5
-3
Two
7
-6
Table 4: Utility schedule, ut , for the infrequent caffeine consumer
Holding without loss of generality an assumption that δ = 1, we can determine what the
caffeine drinker will do. At time zero, the consumer begins to consider the choice but is
not yet at the point of purchase. She evaluates the alternatives as if she were maximizing
Ue . Since Ue equals 0, 2, and 1 upon consuming 0, 1, and 2 shots of espresso, respectively,
the time-zero individual would want the time-one consumer to opt for one shot of espresso.
If the time-one consumer applies the same time-consistent preferences (i.e., acts as if her
βu = 1), her choice will match her desired choice at time zero.
Let us now allow the time-one consumer to have βu 6= 1. Consider what happens when
βu =
1
2
and the consumer pays too little attention to the negative utility in time two. The
decision utility function implies that 0, 1, and 2 shots of espresso will deliver 0, 3.5, and 4
utils of utility, respectively. Since xd = 2, the consumer experiences a utility equal to 1 (i.e.,
Ue (xd = 2) = 1) and an internality equal to 1, since Ue (x∗ = 1) − Ue (xd = 2) = 1. A similar
economic loss can occur if the time-one decision maker acts as if βu = 2 and the consumer
pays too much attention to the negative utility in time two. The decision utility function
equals 0, -1, and -5 for 0, 1, and 2 shots, respectively. She now opts for the zero-shot option.
The consumer experiences a utility equal to 0 (i.e., Ue (xd = 0) = 0) and an internality
equal to 2, since Ue (x∗ = 1) − Ue (xd = 0) = 2. These examples illustrate that a gap
between experienced and decision utility can impose economic costs on the consumer. These
economic costs stem from the wedge between a rational parameter value (i.e., βu = 1) and
that actually used by the consumer in her decision utility function. As demonstrated in these
examples, the internality equals the difference between the maximum level of experienced
utility attainable by a time-consistent consumer and the experienced utility stemming from
the choice made by the time-inconsistent consumer.
Before concluding our example, we highlight an implication of the choice set available to
the consumer. The choice set we considered is discrete, since the consumer cannot choose a
countably infinite number of shots between 0 and 2. This feature implies that for some values
44
βu
xd
Ue (xd )
Ue (x∗ )
Internality
1
One shot
2
2
0
3
4
1
2
3
2
One shot
2
2
0
Two shots
1
2
1
One shot
2
2
0
2
Zero shots
0
2
2
Table 5: Utility measures associated with the choice of zero, one, or two shots of espresso
for a time-one consumer.
of βu sufficiently close to 1, xd = x∗ and the consumer does not experience the economic costs
of internalities. Consider the utility outcomes when βu =
3
4
or βu = 32 . Table 5 summarizes
the experienced and decision utility outcomes and internalities associated with these and the
above βu values. We observe that when βu =
3
4
or βu = 32 , the consumer opts for 1 shot
of espresso. Though wedges between the decision and experienced utility functions exist in
each case, they do not imply a loss in consumer welfare.
This example of espresso decision-making can be immediately extended to the context
of the Energy Star program and appliance choice. When the consumer acts as if she is
optimizing a decision utility function with a βu equal to something other than one, the
consumer makes decisions that she would not make if she were to “rationally” consider the
entire stream of utilities stemming from her decision. In the context of the Energy Star
program, the purchase of lower-priced but less efficient goods delivers immediate benefits,
while the stream of implied energy consumption implies future costs. Consumers undervaluing the costs of energy consumption can be described as acting with a βu < 1 and those
over-valuing, with a βu > 1. Though βu has usually reflected impatience among consumers,
βu < 1 could equally well follow from inattentiveness and βu > 1 from a tendency to be
overattentive.
Moreover, the caffeine example illustrates several concepts that help understand our
subsequent analysis of the Energy Star label and its effects. First, in our assessment of the
Energy Star program, we calculate a normative benchmark for experienced utility that is
analogous to the setting in which βu = 1. Second, larger gaps between the time-consistent
parameter value and that applied by the time-one decision maker imply a greater potential
for the individual to make choices that imply a loss in welfare. Third, because the choice
set is discrete, there are instances in which the use of a time-inconsistent parameter does
45
not imply a loss in welfare to the consumer. Fourth, when such losses occur, they imply real
economic costs to the consumer.
As evident in Section 4.2, our analysis of the Energy Star experiment makes two strong
assumptions. First, we assume that aside from the utility coefficient on the energy consumption attribute, all estimated coefficients reflect standard and thus time consistent preferences.
Since we model utility as additively separable, this is consistent with beta-delta discounting. In our context, the βu parameter applies only to that component of utility stemming
from energy consumption, and any gaps between decision and experienced utility are attributable to either an over- or under-weighing of the implications of the future costs of
energy consumption.
The second assumption is that of the magnitude and duration of energy savings that
obtain for each individual. We were able to quantify internalities in the simple example
only because we could apply a known experienced utility schedule, but such valuation is
frequently quite difficult. This difficulty may at times preclude an economic evaluation
with all relevant factors considered. Nonetheless, for a policy analysis that focuses only
on externalities to be correct, it must make either of two strong assumptions itself. One
assumption is that the Energy Star has no impact on the weights consumers assign to the
energy consumption attribute of alternatives and thus on internalities. The alternative states
that though the Energy Star exerts some influence on internalities, changes in internalities
would be the same no matter which instrument is used. Assumption 2 seems intuitively
unreasonable. The significant Energy Star and Consumption interaction term in Table 1
suggests Assumption 1 is also unjustified.
We note that our discussion of beta-delta preferences provides just one way of conceptualizing mistakes in optimization by consumers. Our simple model here should not be interpreted as a structural representation, as many phenomena could explain an under-investment
in energy efficient goods. For example, Tsvetanov and Segerson (2013) model decisions over
energy-consuming goods as an internal conflict between a fully rational planner and a lessthan-rational doer, in the tradition of Gul and Pesendorfer (2001). The essential element of
any modeling representation is that it admits the possibility for non-optimal decision-making
by consumers. It is this non-optimality that leads to the experienced utility gap we term
the internality.
46
Additional intuition for why policy analysis must condition on choice sets
Figure 10 provides an example of a choice set that may be available to the consumer.
The top curve depicts the total cost of ownership of a bulb, and this is equal to the vertical
sum of the other two curves, which plot the purchase price and lifetime costs of electricity
consumption. A consumer who is over- or under-attentive to costs of electricity consumption
would act as if the lifetime cost curve were shifted upward or downward. Consumer behavior
consistent with the use of a θi 6= θiR in the decision utility function could lead the consumer
to select a bulb with a power rating other than the one that is her experienced utilitymaximizing alternative.
Figure 10: An over- or under-valuation of the costs of electricity consumption is equivalent
to shifting upward or downward the blue curve (the bottom-most curve on the left side of
the graph). This shift would alter the “perceived” total cost of ownership curve such that
the consumer identifies a different power rating as minimizing costs.
By definition, the selection of a such a bulb would imply a loss in experienced utility. Though
we depict in the main text curves relating θi and Ue directly, the relationship is indirect in the
sense that when θi 6= θiR the bulb selected may not be the one that maximizes experienced
utility. Any difference in the experienced utility function stems from subsequent changes in
the bulb selected.
This last point is critically important. If Figure 10 instead included cost and price curves
that implied a total cost of ownership that sharply increased from the true cost-minimizing
47
option, the use of a θi value different from θiR would be unlikely to imply a difference in the
utility experienced by the consumer. Even if the consumer were to apply θi 6= θiR , the selected
bulb is quite likely to be the same as the one that maximizes experienced utility. Thus, the
degree of curvature in the choice set as embedded in the total cost curve is important in
determining the likelihood with which reductions in experienced utility will obtain.
Details on measures collected and sample size
Measures We collected measures that were relevant to either appliance or financial
decision-making, given the similarities in the theorized decision-making processes involved
in both contexts. To ensure that the questions relevant to these measures did not influence
participants’ choice behavior, our queries followed the stated choice exercise. Our measures
can be grouped into four sets: (1) environmental concern and financial status, (2) cognitive
ability, (3) demographics, and (4) psychometrics. We describe the measures included in each
of these buckets below.
Environmental concern and financial status: Consumers may differ in the degree to which
environmental and financial aspects of energy consumption are salient in their decisionmaking processes and thus in the extent to which these concerns are reflected in their utility
coefficient on energy consumption. We measure environmental concern with scores on the
New Ecological Paradigm (NEP) scale (Dunlap et al., 2000). We complemented this measure
with questions about environmental self-efficacy, or the degree to which the participant
believed his decisions could impact environmental outcomes. We do not use the latter here
because the measure uses questions primarily relevant to refrigerator choice.
To characterize financial status, we elicited temporal discount rates, household assets
and liabilities, and liquidity constraints. Our estimates of discount rates use the Kirby
and Marakovic (1996) methodology.47 We used these discount rates to establish the rational
benchmark coefficient on energy consumption. Asset and liability measures allow us to gauge
47
Each participant answered 21 questions in which they were asked to choose either a smaller immediate
payoff or a larger delayed payoff. When ordered, the questions imply 22 exponential and hyperbolic discount
rates that predict a switch from the immediate payoff to the delayed payoff. The original paper includes
21 values, but we add two more bounds, 0 and 1, that allow participants to always select the immediate
or delayed payoff. For each participant, we calculate the predicted choices based on each of the 22 time
preferences and assign to the subject the discount rate that maximizes the match between actual and
predicted choices. We treated accuracy ties between rates as in Kirby and Marakovic (1996).
48
the wealth of each household, and liquidity constraints, to understand the degree to which
this one factor contributes to price sensitivity in consumers.
We include two other measures that have provided a richer description of financial decision making. A metric of self-continuity describes the degree to which one identifies with
one’s future selves. Our measure is based on that used by Ersner-Hershfield et al. (2009),
who show that differences in self-continuity can explain differences in saving rates. This corresponds to the βu parameter from Section 4.1 and, in the context of the utility coefficient
on energy consumption, to over- or under-valuation of savings on energy consumption. The
metric could thus explain differences in the valuation of savings in energy consumption which
accrue to future selves. The other metric measures the emotional states of our participants.
Emotional affect has implications for belief formation and financial decision-making: positive emotions such as excitement induce risk taking and confidence, while negative emotions
trigger opposite reactions (Kuhnen and Knutson, 2011). Decision-making involving energy
consumption entails risks about future electricity prices and environmental damages associated with energy consumption. We use responses from the Positive and Negative Affect
Schedule (PANAS) (Watson, Clark, and Tellegan, 1988) to measure the degree to which
individuals are predisposed toward positive or negative affect.
Eight questions were intended to examine risk preferences among participants. One set
asked how individuals would allocate a sudden windfall of $10,000 among stocks, bonds, and
cash. Another asked how their current assets were allocated. A final measure sought to
measure a gap between consumers’ willingness to pay to take and to avoid a gamble. We do
not use the first two measures, which are part of a survey in development but not definitively
linked to risk preferences. We do not use the questions about willingness to pay because the
answers we received indicated a misunderstanding of the questions.
Cognitive ability: Behavior consistent with the maximization of a time-consistent experienced utility function requires the ability to focus on all relevant information and translate
this into utility implications. We include two measures to characterize participants’ abilities
to perform these steps. One is the cognitive reflection test (CRT) that determines the degree to which the individual performs slow and deliberate calculations instead of quick and
reflexive ones, and we use a three-item measure developed by Frederick (2005). A separate
numeracy measure is based on questions developed by Peters et al. (2007). We supplement
the measure with an additional question that tests’ individuals understanding of conditional
49
probabilities.
Demographics: Our data include place of birth and residence, age, household size, household income, energy bill burden, home ownership status, subjective socioeconomic status,
education, intra-household income distribution, and expected length of stay in current residence. We also asked whether the respondent was responsible for financial and appliance
decisions. Our education metric includes four buckets. The first includes those with at most
a 2 year college degree, the second, those with a 4 year college degree, the third, those with
a professional degree (e.g., an MBA or JD), and the fourth, those with a non-professional
masters, doctoral degree, or medical degree.
Psychometrics: We measure the degree to which each participant is characterized by each
of the Big-Five personality domains, namely, extraversion, agreeableness, conscientiousness,
emotional stability, and openness to new experiences. Our inclusion of the measures is
driven largely by an expectation that differences in conscientiousness could help explain
attentiveness to energy consumption information and the utilization of the Energy Star
logo. We use a ten item inventory developed by Gosling, Rentfrow, and Swann (2003).
Sample size We based our target sample size of 1,500 on the results of a pilot phase of
data collection. For our pilot, we recruited participants via the Amazon Mechanical Turk
platform and gathered data using the Qualtrics survey tool. We randomized participants
into one of 8 blocks of 16 choice sets. Each participant selected one of three alternate light
bulbs in 16 light bulb choice sets. We used Ngene (ChoiceMetrics, 2012) to generate the 128
choice sets included in the experiment. Of the 110 submitted responses, we deemed 103 to
be complete and attentive by checking whether participants responded correctly to questions
about their online identity in the middle and at the end of the task. We used a subset of
92 individuals in the upper 90 percentiles of task completion time to estimate coefficients.
All participants were aged 18 or over, U.S. residents, and registered users of the Amazon
Mechanical Turk system. We provided each participant $7. We derived priors for our main
choice experiment by estimating coefficients of a multinomial logit model of the pilot choice
data. Among the coefficients we estimated was one on a binary variable that coded whether
an alternative light bulb was labeled with the Energy Star.
Our initial objective for our main experiment was to understand whether the presence
of the Energy Star was a significant predictor of choice and whether the magnitude of the
Energy Star’s effect differed by the three experimental conditions. We used the estimated
50
coefficient and standard error on the binary variable that coded for the presence of the
Energy Star label to calculate the sample size requirement.
Our main experiment reflected several changes to the pilot data set. First, our pilot data
set included two types of light bulb technologies (i.e., CFL and light emitting diodes (LED)),
while our main experiment used just CFL bulbs. Second, our utility specification in the main
experiment was richer than that in the pilot setting. Third, and finally, we planned to use a
mixed logit formulation to model the choice data from the main experiment instead of the
fixed coefficient multinomial logit approach we used with the pilot data. Given all of these
changes, we scaled up our sample size as permitted by the study’s budgetary constraints.
51
52
$60 - 80K
Numeracy (Number correct of 15)
10.16
9.82
4.88
Big Five-Emotion Stability (min: 2; max: 14)
Big Five-Openness to Exp. (min: 2; max: 14)
Future Self-Similarity (min: 1; max: 7)
4.94
9.91
10.14
11.42
10.47
8.08
17.9
32.6
51.3
5
0.79
0.023
4.94
9.80
10.12
11.31
10.51
8.24
17.8
32.5
50.9
5
$60 - 80K
10.21
0.74
0.017
0.022
80.2
2.30
2.79
52.2%
50.3
529
Cond. 3
4.92
9.84
10.14
11.29
10.49
8.26
18.2
32.8
51.0
5
$60 - 80K
10.27
0.75
0.020
0.027
79.4
2.32
2.76
50.0%
50.3
1550
Total
0.7
0.3
0.06
0.09
0.02
0.28
0.06
0.02
0.04
0.17
0.02
0.38
0.22∗
0.02
0.9
0.9
0.2
–
–
0.13
0.02
0.004
0.005
1.8
0.00
0.05
2.8%
0.8
0.8
0.6
–
–
0.08
0.07
0.002
0.004
1.4
0.06
0.07
1.1%
15
(1/3)
(1/2)
7
Diff
Diff
0.00
0.11
0.02
0.11
0.04
0.16
0.1
0.1
0.4
–
–
0.05
0.05
0.006
0.009
0.4
0.06
0.02
3.9%
0.4
22
(2/3)
Diff
three conditions.
similarly across experimental conditions; e.g., the distributions of financial and appliance decision-making are similar across the
significance label are marked by an asterisk. Some variables are not summarized above, but these are factors that are distributed
Table 6: Overview of demographic, psychological and financial metrics across the sample. Significant differences at the 0.05
10.49
11.14
Big Five-Extraversion (min: 2; max: 14)
Big Five-Conscientious (min: 2; max: 14)
8.46
PANAS-Negative (min: 10; max: 50)
PANAS-Positive (min: 10; max: 50)
Big Five-Agreeable (min: 2; max: 14)
50.7
18.7
33.4
NEP (min: 15; max: 75)
Socio-economic Status (Median)
5
$60 - 80K
0.72
10.34
CRT Score (Number correct of 3)
Household Income (Median)
10.26
0.021
Exponential Discount Rate
0.031
79.8
2.36
78.4
Education (of 4)
2.77
0.027
2.30
Household Size
49.9
48.3%
Hyperbolic Discount Rate
2.84
% Women
507
Cond. 2
Dollar + Energy Dollar
Units
Units
Primary Earner, Wage Share (%)
50.6
49.4%
Age
514
Cond. 1
Energy
Units
Participants
Attribute
References
Allcott, H. (2011), “Consumers’ perceptions and misperceptions of energy costs,” The American Economic Review, 101(3), 98–104.
Allcott, H., S. Mullainathan, and D. Taubinsky (2014), “Energy policy with externalities
and internalities,” Journal of Public Economics, 112, 72–88.
Allenby, G. M., and P. E. Rossi (1998), “Marketing models of consumer heterogeneity,”
Journal of Econometrics, 89(1), 57–78.
Chetty, R., A. Looney, and K. Kroft (2009), “Salience and Taxation: Theory and Evidence,”
The American Economic Review, 99(4), 1145 – 1177.
ChoiceMetrics (2012), “Ngene,” Version 1.1.1.
Daly, A., S. Hess, and K. Train (2012), “Assuring finite moments for willingness to pay in
random coefficient models,” Transportation, 39(1), 19–31.
Dunlap, R., K. V. Liere, A. Mertig, and R. Jones (2000), “Measuring Endorsement of the
New Ecological Paradigm: A Revised NEP Scale,” Journal of Social Issues, 56(3), 425 –
442.
Ersner-Hershfield, H., M. Garton, K. Ballard, G. Samanez-Larkin, and B. Knutson (2009),
“Don’t stop thinking about tomorrow: Individual differences in future self-continuity account for saving,” Judgment and Decision Making, 4(4), 280 – 286.
Frederick, S. (2005), “Cognitive Reflection and Decision Making,” Journal of Economic
Perspectives, 19(4), 25 – 42.
Gillingham, K., and K. Palmer (2014), “Bridging the energy efficiency gap: Policy insights
from economic theory and empirical evidence,” Review of Environmental Economics and
Policy, 8(1), 18–38.
Gosling, S., P. Rentfrow, and W. Swann (2003), “A very brief measure of the Big-Five
personality domains,” Journal of Research in Personality, 37(6), 504 – 528.
Gul, F., and W. Pesendorfer (2001), “Temptation and self-control,” Econometrica, 69(6),
1403–1435.
53
Hausman, J. A. (1979), “Individual discount rates and the purchase and utilization of energyusing durables,” Bell Journal of Economics, 10(1), 33–54.
Hensher, D., and W. Greene (2003), “The Mixed Logit Model: The State of Practice,”
Transportation, 30(2), 133 – 176.
Herrnstein, R. J., G. F. Loewenstein, D. Prelec, and W. Vaughan (1993), “Utility maximization and melioration: Internalities in individual choice,” Journal of Behavioral Decision
Making, 6(3), 149–185.
Hole, A. R. (2007), “Estimating mixed logit models using maximum simulated likelihood,”
Stata Journal, 7(3), 388–401.
Houde, S. (2013), “Bunching with the Stars: How Firms Respond to Environmental Certification,” Mimeo, University of Maryland.
Houde, S. (2014), “How Consumers Respond to Environmental Certification and the Value
of Energy Information,” NBER WP No. 20019.
Kirby, K., and N. Marakovic (1996), “Delay-discounting probabilistic rewards: Rates decrease as amounts increase,” Psychonomic Bulletin and Review, 3(1), 100 – 104.
Kuhnen, C., and B. Knutson (2011), “The influence of Affect on Beliefs, Preferences, and
Financial Decisions,” Journal of Financial and Quantitative Analysis, 46(3), 605 – 626.
Laibson, D. (1997), “Golden eggs and hyperbolic discounting,” The Quarterly Journal of
Economics, 112(2), 443 – 478.
Leard, B. (2013), “Consumer Heterogeneity and the Energy Paradox,” Mimeo, Cornell University.
Min, J., I. L. Azevedo, J. Michalek, and W. B. de Bruin (2014), “Labeling energy cost on
light bulbs lowers implicit discount rates,” Ecological Economics, 97, 42–50.
Newell, R. G., and J. Siikamaki (2014), “Nudging Energy Efficiency Behavior: The Role of Information Labels,” Journal of the Association of Environmental and Resource Economists,
1(4), 555–598.
54
O’Donoghue, T., and M. Rabin (2000), “The economics of immediate gratification,” Journal
of Behavioral Decision Making, 13(2), 233–250.
Peters, E., N. Dieckmann, A. Dixon, J. H. Hibbard, and C. Mertz (2007), “Less is more in
presenting quality information to consumers,” Medical Care Research and Review, 64(2),
169–190.
Revelt, D., and K. Train (1998), “Mixed logit with repeated choices: households’ choices of
appliance efficiency level,” Review of Economics and Statistics, 80(4), 647–657.
Sallee, J. M. (2014), “Rational Inattention and Energy Efficiency,” Journal of Law and
Economics, 57(3), 781–820.
Sammer, K., and R. Wustenhagen (2006), “The influence of eco-labelling on consumer behaviourResults of a discrete choice analysis for washing machines,” Business Strategy and
the Environment, 15(3), 185 – 199.
Shen, J., and T. Saijo (2009), “Does an energy efficiency label alter consumers’ purchasing
decisions? A latent class approach based on a stated choice experiment in Shanghai,”
Journal of Environmental Management, 90(11), 3561 – 3573.
Small, K. A., and H. S. Rosen (1981), “Applied Welfare Economics with Discrete Choice
Models,” Econometrica, 105–130.
Sundqvist, T. (2004), “What causes the disparity of electricity externality estimates,” Energy
Policy, 1753–1766.
Swait, J., and J. Louviere (1993), “The role of the scale parameter in the estimation and
comparison of multinomial logit models,” Journal of Marketing Research, 305–314.
Thaler, R., and C. Sunstein (2008), Nudge: Improving Decisions About Health, Wealth, and
Happiness, Penguin.
Train, K. (2009), Discrete Choice Methods with Simulation, Cambridge University Press.
Tsvetanov, T., and K. Segerson (2013), “Re-evaluating the role of energy efficiency standards:
A behavioral economics approach,” Journal of Environmental Economics and Management, 66(2), 347–363.
55
U.S. Census Bureau (2014), “Residential Vacancies and Homeownership in the Fourth Quarter 2013,” http://www.census.gov/housing/hvs/files/qtr413/q413press.pdf.
Ward, D., C. Clark, K. Jensen, S. Yen, and C. Russell (2011), “Factors influencing
willingness-to-pay for the ENERGY STAR label,” Energy Policy, 39(3), 1450 – 1458.
Watson, D., L. Clark, and A. Tellegan (1988), “Development and validation of brief measures
of positive and negative affect: The PANAS scales,” Journal of Personality and Social
Psychology, 54(6), 1063 – 1070.
Wooldridge, J. M. (2010), Econometric Analysis of Cross Section and Panel Data, MIT
press.
56